Incrementality Testing for AI-Managed Ad Campaigns: How to Check a Platform’s Lift Claim

When an ad platform hands campaign decisions to an AI agent, it also reports the results. The only reliable way to know whether the agent caused extra sales is to compare exposed customers with a comparable group that never saw the ads.

Why platform lift needs a second opinion

Ad platforms now run more of the campaign themselves. On September 29, 2026, Amazon Ads renamed its unified buying platform Amazon Ads Agent and described it as powered by agentic AI, with Full-Funnel Campaigns available to all US advertisers.

The announcement also reported that advertisers who applied its natural-language targeting recommendations saw “more than 25% additional unique customers” and more than 10% lower cost per impression on average. These are Amazon’s own figures. We found no sample size or control design published alongside them.

That does not make the claim false. It means the number answers a vendor’s question, which is whether the feature is worth promoting, and not yours, which is whether your next dollar produced a sale that would not have happened anyway. Incrementality, meaning the extra conversions caused by advertising, is the measure that separates those two questions.

What the research says about weak comparisons

The common shortcut is to compare people who saw an ad with people who did not, then adjust for age, interests and past behavior. Research suggests this often fails. Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky tested it in a paper published in Marketing Science in 2019.

They used 15 US advertising experiments at Facebook, covering 500 million user-experiment observations and 1.6 billion ad impressions, and compared the randomized results with several observational models. The observational methods often failed to match the experiments, even after conditioning on extensive demographic and behavioral variables.

The authors also note that advertising’s effect on outcomes tends to be small, which makes bias easy to introduce. For a buyer, the implication is practical: a dashboard comparison of exposed and unexposed customers is a hypothesis, and a randomized or geographic holdout is the test. The same caution applies to the attribution disagreements we covered in why attribution models disagree.

Three ways to build a control group

A control group is the set of people or places deliberately kept from seeing the campaign. Google’s help page on Conversion Lift defines lift as “the increase in conversions that are caused by the presence of the ad” and describes a treatment group that sees ads and a control group that does not.

Method	How the control is formed	Main constraint
Platform user-based study	The platform splits eligible users into exposed and held-out groups	Google says Conversion Lift is available only for some accounts, via a representative
Geography-based study	Matched regions get the campaign, others do not	Needs enough regional volume to detect a small effect
Campaign split or pause test	Spend is switched off for a set period or region, then results are compared	Seasonality and promotions can distort the comparison

Google’s page describes both user-based and geography-based Conversion Lift, and notes that the geography type supports offline data. That matters for lead-generation firms whose sales close by phone or in a CRM, because the conversion often happens outside the ad platform.

Running a test, step by step

The order matters more than the tool. Decisions made after seeing results are the main way tests mislead.

1. Write the decision the test will inform, such as whether to move budget to the agent-run campaign.
2. Choose one outcome that reflects revenue, such as qualified leads or purchases, not clicks.
3. Fix the control group, whether users or regions, before launch.
4. Keep budget, creative and offers steady in both groups for the whole period.
5. Set the run length in advance so it covers your normal buying cycle.
6. Compare conversions between groups, divide extra conversions into spend, and decide.

Step four is where AI-managed campaigns create trouble. If the agent shifts targeting daily, the exposed group changes shape mid-test. Ask the vendor how the holdout stays fixed, and see our post on evaluating an agentic marketing vendor for related questions.

When a test cannot answer the question

Small accounts often cannot detect small effects. Because advertising effects tend to be small, as Gordon and colleagues note, a low-volume test can return a result that looks like zero lift when the sample was simply too small to see one. An inconclusive test is not evidence of no effect.

In that case, a longer test, a larger region split or a bigger budget increase is a better response than trusting the platform dashboard. A test also measures one period. Results from a holiday quarter may not hold in January.

Pairing a test with human review is sensible while you build confidence. Our piece on approval gates for agentic ad spend covers how to limit what an agent can change, and our 90-day pilot guide shows where a holdout fits inside it.

Frequently Asked Questions

What is incrementality in advertising?

It is the number of conversions caused by the ads, measured as the difference between people who could see them and a comparable group who could not. Conversions that would have happened anyway do not count.

Can a small business run an incrementality test?

Yes, but the test may be too small to detect modest effects. A geography split or a planned pause on one campaign is usually the most practical option, and results should be read as indicative.

Should I trust the lift numbers an AI ad platform reports?

Treat them as vendor claims until you have a holdout result from your own account. The figures may be accurate in aggregate and still not describe your campaigns.

As ad platforms take on more decisions, the buyer’s job shifts toward proving what those decisions cause. At Grow with Hafsa we build performance-marketing and AI agent programs around a pre-agreed test and baseline, and we are happy to look at how a holdout could work for the campaigns you already run.

Google’s October 2026 AI Mode Link Changes: What Was Announced and What Is Still Unmeasured

On October 15, 2026, Google said it would show more links, higher up, inside AI Mode and AI Overviews. For a business that depends on organic leads, the useful reading is that the announcement changes where links appear, but it publishes no evidence that clicks will rise.

What Google announced

Google published “More ways to connect with the web from Search” on October 15, 2026, written by Hema Budaraju, a vice president of product management. Search Engine Journal and PPC Land both summarized it the same day or shortly after, and their accounts agree on the five changes.

The first is that inline links, meaning links placed beside the sentence they support, now appear higher and in more responses. Google described “continued and substantial increases” but gave no figure. The other four changes are smaller and uneven in status, as the table shows.

Change	Status Google gave	What the post leaves unstated
Inline links higher in responses	Rolling out	How much higher, and how often
Official-site section at the top for task-based queries	Test	Start date, markets, selection criteria
Image thumbnails on links at the end of responses	Design change	Which responses get them
Timely-article carousels for Preferred Sources	Expanding	Which locations and languages
Clickable creator names on AI Mode video results	Test, started the week of October 12	Markets (PPC Land infers the US, Google did not say)

The creator feature links to Search profiles, which Search Engine Journal reports are limited to the US, Canada and India and to accounts with at least 10,000 followers on YouTube, Instagram, X or TikTok. Most B2B service firms will not qualify, so it matters less than the inline-link change.

What the post does not measure

The central claim is that AI answers now send people to the open web more often. The evidence offered is thin. PPC Land’s review of the post found no click-through data, and the one growth figure, that Preferred Sources use rose “by more than 7x since the start of the year,” has no January baseline and no definition of “use.”

The same article notes that earlier Preferred Sources milestones counted unique sources selected, not users: about 90,000 on December 10, 2025, more than 345,000 on May 27, 2026, and more than 600,000 on August 20, 2026. Those numbers describe publisher selection, which is a different thing from reader behavior.

The implication for a buyer of SEO or GEO (Generative Engine Optimization) services is direct. A vendor who cites this announcement as proof that AI search traffic will recover is repeating Google’s framing, not a measured result. Independent click studies, covered in our review of what Pew and Ahrefs measured, still describe fewer clicks where AI summaries appear.

Why Search Console will not show the effect

You cannot easily verify the change on your own data. Search Engine Journal reports that Search Console gives every link inside an AI Overview the same single position, so an inline link that moves higher in the answer does not change its reported position.

Google’s own guidance on AI features in Search says sites appearing in them are “included in the overall search traffic in Search Console” and reported under the Web search type. There is no separate AI Mode filter described on that page. Traffic from these surfaces is therefore blended with ordinary results.

In practice, you will see movement only as a change in clicks to specific landing pages, and you will not be able to attribute it cleanly to this announcement. That is why a dated baseline matters more than any new tool. Our guide to measuring AI search visibility when answers change every run covers the sampling approach that fills part of this gap.

What to do this week

None of these steps depend on Google’s tests succeeding. They are low-cost and hold up whether or not the links drive more visits.

1. Export 90 days of Search Console clicks and impressions for your ten most commercial pages, and note October 15, 2026 as the reference date.
2. Confirm those pages are indexed and eligible to show with a snippet. Google states a page needs this to appear as a supporting link, adds that there are no additional technical requirements, and cautions that appearance is not guaranteed.
3. Make sure the pages a buyer would need for a task, such as pricing, booking and contact, are clear and crawlable, since the official-site test targets task-based queries. This is our inference from the test’s description, not a Google instruction.
4. Check your referral reports for AI sources monthly, using the approach in our post on tracking AI referral traffic.
5. Hold any budget change until two or more months of your own data show a pattern.

Preferred Sources deserves one extra check. Search Engine Journal reports that qualifying requires the site to be included in Search generative AI features in Search Console, so confirm that setting rather than assuming it.

Frequently Asked Questions

Did Google say AI Mode will send more traffic to websites?

No. Google said it is making links more prominent and would refine the experience based on what works best. The post does not include click-through data or a traffic forecast.

Can I see AI Mode clicks separately in Search Console?

Not according to Google’s guidance, which says AI feature traffic is counted within overall search traffic under the Web search type. Search Console also gives all links in an AI Overview one position.

Does this change what GEO work is worth doing?

It does not change the fundamentals. Indexable, clearly written pages that answer buyer questions remain the base, and the link changes only affect how visibly those pages are cited.

Announcements like this one are worth tracking but rarely worth reacting to in isolation. At Grow with Hafsa we help teams set up the baselines and measurement that turn a platform change into a decision, as part of our AI marketing and performance-marketing work, and we are glad to review how your current reporting would catch it.

How to Pilot an AI Marketing Agent: A 90-Day Scope With Exit Criteria

Signing a twelve-month agentic marketing contract before seeing the agent work on your account is the most expensive way to learn. A 90-day pilot with one workflow, a baseline and written exit criteria gives you evidence a sales demo cannot.

Why pilot before committing

Gartner’s June 25, 2025 press release predicted that over 40% of agentic AI projects will be canceled by the end of 2027. It attributed the cancellations to escalating costs, unclear business value or inadequate risk controls.

The same release described “agent washing”, where vendors rebrand existing products such as chatbots, assistants and robotic process automation without real agentic capability. Gartner estimated that only about 130 of the thousands of vendors claiming agentic features are genuine.

For a buyer, the implication is practical. The chance that a given offer is oversold is not small, and the cost of finding out should be capped. A pilot turns the vendor’s claims into a test on your own data, with a defined end. As we covered in our guide to evaluating an agentic marketing vendor, even a survey of 250 agencies found self-reported returns overstated, so your own measurements matter more than any benchmark.

Choosing the workflow

Pilot workflow: a single repeatable task the agent performs end to end, small enough to measure and safe enough to reverse. Starting with several workflows blurs which one produced a result.

Three tests help you choose. The task should be reversible, meaning a mistake can be undone without public damage. It should be measurable against a number you already track. And it should be bounded, so the agent cannot touch budgets, live pages or customer messages without approval.

Teams often pick the most impressive workflow, such as fully automated publishing. We recommend the opposite: pick the one where a wrong output costs you an hour, not a ranking or a refund. Our overview of signs your business is ready for AI agent workflows covers the readiness checks to run first.

Good and poor pilot candidates

The table applies those three tests to common agentic SEO and ads tasks. It is our judgment, offered as a starting point you should adjust to your own risk tolerance.

Task	Reversible	Measurable	Pilot fit
Technical SEO audit reports for human review	Yes, nothing changes live	Issues found versus a known list	Strong
Content refresh briefs for existing pages	Yes, drafts only	Editor acceptance rate	Strong
AI search visibility monitoring	Yes, read-only	Consistency across repeated runs	Strong
Auto-publishing new pages	Partly, but indexing is hard to undo	Rankings, slow to show	Poor at first
Changing ad budgets or bids unattended	No, spend is gone	Fast, but risky	Poor without approval gates

For the last row, our piece on approval gates for agentic ad spend explains why spend should keep a human checkpoint.

The 90-day plan

Each phase has one purpose, and you move on only when its output exists on paper. This is a working structure we suggest, not a validated method.

1. Weeks 1 and 2: record the baseline for your chosen metric, agree the cost cap, and build a small test set of past cases with known correct answers.
2. Weeks 3 to 6: run the agent in shadow mode, where it produces outputs but nothing goes live, and score each against the test set.
3. Weeks 7 to 10: allow supervised live use, with a named person approving every output that reaches customers, search engines or ad accounts.
4. Weeks 11 and 12: compare results with the baseline, review costs and failures, and make the decision against the criteria set in week 1.

Shadow mode is the phase most teams skip. It is also the cheapest way to find failures, because mistakes cost nothing yet.

Writing exit criteria

Exit criteria: thresholds agreed in writing before the pilot starts that decide whether to continue, adjust or stop. They protect you from judging results after the fact, when sunk cost and enthusiasm distort the call.

Useful criteria are specific. For example: the agent must match or beat the human baseline on accuracy in the test set, total usage and vendor fees must stay under the cap, and no output may reach a customer or ad account without approval.

Add a stop rule for serious failures, such as any change that breaks a live page or overspends a budget. Gartner named unclear business value and inadequate risk controls among the reasons projects fail, and these two clauses address both directly. If the vendor will not agree to written criteria, that tells you something useful before you have spent anything.

What a pilot cannot prove

A 90-day window is short for search. Rankings and AI citations can take months to move, so judge an SEO pilot on process quality, accuracy and time saved, not on traffic gains.

Small pilots also hide scaling costs. Usage fees, review time and error rates can change when volume grows tenfold, so ask the vendor to model them before you expand. Finally, a pilot on one workflow says little about the vendor’s other claims. Treat each new workflow as its own, smaller test.

Frequently Asked Questions

How long should an AI marketing pilot last?

Ninety days is a practical window: two weeks of setup, four of shadow testing, four of supervised use and two to decide. This is our suggestion, not an industry standard.

What should I measure in an agentic SEO pilot?

Measure accuracy against a test set, reviewer acceptance rate, hours saved and total cost. Rankings and traffic move too slowly to judge within 90 days.

Should a pilot include ad budget changes?

Not at first. Spend cannot be recovered, so keep a human approval step on any budget or bid change until the agent has a clean record on your account.

A pilot is easier to run with someone who has set one up before. At Grow with Hafsa we scope AI agent and performance-marketing pilots with baselines and approval gates built in, and we are happy to review a plan you are drafting.

A pilot is also the natural place for a holdout; incrementality testing for AI-managed ad campaigns explains how to build one.

AI Overviews on Commercial Searches: What Semrush’s Six-Month Study Means for Buyer-Intent Pages

Most buyers do not search the way they used to. The Semrush data suggests Google’s AI answers are moving into the research stage of commercial searches, which is exactly where service businesses win or lose a shortlist spot.

What the study measured

Semrush published its analysis on July 2, 2026. It tracked more than 600,000 keywords from its US desktop database across 10 industries, from November 2025 to April 2026, and segmented them by industry, search intent and cost per click tier.

The headline result is a relative change. AI Overviews, the AI-written summaries Google places above organic results, grew an average of 71% across results pages with commercial intent. Every industry in the sample saw growth, though at very different rates.

Finance rose 231.25%, computers and electronics 107.62%, and games 76.66%. Because these are growth rates, they do not tell you what share of searches show an AI Overview today. They tell you the direction and speed of change. For a buyer-facing business, the direction is the useful part: more of the research journey now passes through an AI answer before a click happens.

Why commercial intent is the exposed stage

Commercial intent: a search by someone researching options before choosing one, such as comparing providers or reading reviews. Transactional intent: a search by someone ready to act, such as buying or booking.

Semrush found AI Overviews on transactional queries decreased 5%, with most industries flat or declining. Computers and electronics, up 80.45%, and news, up 45.96%, were the exceptions. In other words, Google appears to be adding AI answers where people compare, and holding back where people buy.

That matters for companies that sell considered services. A founder choosing an SEO or ads partner spends most of the journey in commercial research. If an AI summary frames that comparison, your pages compete to be cited in it, not just to rank beneath it. Our analysis of AI Overviews and click-through rate covers what the click data shows once an overview appears.

What changes by intent type

The table pairs Semrush’s findings with our reading of what each means for content planning. The right-hand column is interpretation, not a finding from the study.

Intent	What Semrush found	Our reading for content planning
Commercial	AI Overviews up an average of 71%	Comparison and evaluation pages need to be quotable, with clear claims and named evidence
Transactional	AI Overviews down 5% on average	Product, pricing and booking pages still depend mainly on classic ranking and conversion work
Informational and navigational	Change not reported in the article	Do not assume either way; check your own data

The gap in the last row is worth noting. The Semrush article does not report percentage changes for those two intents, so any claim about them needs another source.

The paid search angle

The study also looked at advertising. Semrush found Google Ads and AI Overviews appear on the same results page roughly twice as often as a year earlier, though the overlap declined in early 2026 while the overall trend stayed upward.

Cost per click, the price paid for each ad click, was higher for keywords with an AI Overview in most industries. In Jobs and education the average was $5.02 with an overview and $1.51 without. In finance it was $4.84 against $2.14.

The pattern was not universal. Keywords with an overview had lower average costs in internet and telecom, games, hobbies and leisure, food and drink, news, and sports. Semrush also notes that Google Ads does not reveal which paid queries trigger an overview, which limits how far advertisers can verify this inside their own accounts. For spend decisions, see our look at ads inside AI answers.

How to audit your own keywords

Industry averages hide your specific exposure. This sequence uses tools most teams already have and takes a few hours.

1. Export the 50 to 100 queries that bring you leads, not just traffic.
2. Label each as commercial, transactional or informational using the definitions above.
3. Search each commercial query in a logged-out browser and note whether an AI Overview appears and who it cites.
4. Compare lead volume from those queries across the last two quarters.
5. Rewrite the three most exposed pages so each key claim is stated plainly, with its source named in the same sentence.

Because overviews vary by location and device, repeat checks over several days. Our guide to measuring AI search visibility when answers change explains why a single check misleads.

What the study cannot tell you

Treat the numbers as directional. The sample is US desktop only and covers six months, so mobile behavior and other countries, including the UK, are not represented.

Semrush offers explanations for why commercial queries are growing, such as longer research stages, but hedges them as possibilities. They were not tested. The company also sells the tools used in the analysis, which is an interest worth keeping in mind.

Finally, the study measures how often an overview appears, not whether it changes lead quality or revenue. Those answers sit in your own analytics, which is why the audit above matters more than the headline percentage.

Frequently Asked Questions

Do AI Overviews appear more on commercial or transactional searches?

In Semrush’s six-month US desktop study, they grew 71% on average for commercial queries and fell 5% for transactional ones, so growth is concentrated in the research stage.

Do AI Overviews raise cost per click?

In most industries Semrush examined, keywords with an overview had higher average costs, but several industries showed lower costs, so check your own account data.

Should a service business rewrite its commercial pages for AI answers?

Start with the pages that drive leads and sit on exposed queries. Make claims specific, attribute evidence inline and keep the information easy to quote.

Deciding where to put effort between organic, paid and AI visibility is easier with your own query data in front of you. At Grow with Hafsa we help teams run this kind of audit as part of AI marketing and performance-marketing work, and we are glad to look at a keyword set you are already weighing.

How to Evaluate an Agentic Marketing Vendor: Ask for the Evaluation Harness First

Every agency now says it runs AI agents. The useful question for a buyer is not whether a vendor uses agents, but how it proves the agents work before they touch your budget, your site or your brand.

What the survey found

Digital Applied surveyed 250 independent marketing and development agencies across the US, EU and APAC between January 14 and March 7, 2026. Agencies had between 8 and 180 staff. The survey reported that 41% had at least one agent in production, up from 9% in the first quarter of 2025, while 58% were still piloting.

Returns were uneven. Median reported ROI was 3.2 times, the top decile reported 11 times, and the bottom quartile reported 0.7 times, which is below break-even. Median token spend, meaning the usage fees paid to AI model providers, was $1,800 a month per active agent.

For a buyer, the spread matters more than the median. Agents are not reliably profitable for the agencies running them, so a vendor’s confidence is not evidence. Your question is whether this vendor sits near the top or the bottom of that range.

What an evaluation harness is

Evaluation harness: a fixed set of test cases with known good outcomes that an agent is run against before launch and after every change. It turns “the agent seems fine” into a measurable pass rate.

The survey’s publisher reported that bottom-quartile agencies almost always lacked a workflow-level harness, and that agencies which built evaluations first reported ROI around 11 times. The report gives no correlation coefficient or subgroup sizes, so this is an association and not a proven cause. It also named evaluation and testing complexity as the top blocker, cited in the top three by 49% of agencies.

In plain terms, a harness for an SEO agent might hold 50 past pages with known issues and check whether the agent finds them without inventing new ones. If a vendor cannot describe its version of this, it is probably judging results by eye.

Weak and strong vendor answers

The table contrasts typical answers to the checks that matter most. It is a judgment aid built from the survey’s themes, not a scored standard.

Check	Weak answer	Strong answer
Testing	“We review outputs as we go”	A named test set, a pass rate and a re-test after each change
Human approval	Agent publishes or spends automatically	Defined approval gates for spend, publishing and client-facing changes
Cost	One flat fee, usage not mentioned	Cost per task and a monthly usage cap
Reporting	Activity counts such as pages produced	Business outcomes tied to your baseline
Explainability	“The model decided”	A log showing what the agent did and why

Explainability is not a side issue. Client trust and explainability was the second most cited blocker in the survey, named by 37% of agencies, and cost predictability the third, at 32%.

Questions to ask, in order

Use these in a first call. Each one is designed to be hard to answer with a slide deck.

1. What tasks will agents perform for us, and what will a human still approve?
2. Show us the test set used to check that work, and the latest pass rate.
3. What does one task cost in usage fees, and what is the monthly cap?
4. What did the last agent failure look like, and how was it caught?
5. How will you measure success against our own baseline, not an industry average?

Approval gates deserve special attention for paid media. Our analysis of approval gates for agentic ad spend explains why spend should never be fully unattended. For the wider vendor-selection picture, see how to choose an AI marketing agency.

How much weight the survey deserves

Treat the numbers as indicative. The sample was not random: recruitment came from the publisher’s own database, a partner list and cold outreach, and the publisher says it over-indexes on agencies already considering agentic AI, so true adoption may be 5 to 10 points lower.

The ROI figures are also self-reported. The publisher compared billing data for 47 agencies and found self-reports overstated ROI by about 18% and understated token spend by about 24%. That is a useful warning for buyers: ask vendors for billed numbers, not claimed ones. The same logic applies to our guide on what agentic SEO actually is, where autonomy claims need the same scrutiny.

Frequently Asked Questions

What is an evaluation harness in AI marketing?

It is a fixed set of test cases with known correct outcomes, used to score an AI agent before launch and after each change, so quality is measured instead of assumed.

Is the 3.2 times ROI figure a reliable benchmark?

Only as a rough guide. It is a self-reported median from a non-random survey of 250 agencies, and the publisher found its own sample overstated ROI by about 18% where billing data existed.

Should a vendor’s agents ever act without human approval?

Low-risk, reversible tasks can run unattended. Spending, publishing and client-facing changes should keep a human approval step until the vendor shows a track record on your account.

Choosing a vendor is easier when you ask for proof of testing and not a demo. At Grow with Hafsa we design AI agent and performance-marketing work with clear approval gates and measurable baselines, and we are happy to pressure-test a proposal you are already weighing.

Once you have chosen a vendor, test it before committing: our guide to piloting an AI marketing agent over 90 days sets out the scope and exit criteria.

Local Business AI Search Visibility: What Two 2026 Studies Say Drives Recommendations

A customer asks an AI assistant for the best plumber, dentist or accountant nearby, and gets a short list of names. Two 2026 studies suggest that list is far shorter than Google’s local results, and that the signals behind it differ from classic technical SEO.

How much harder AI recommendations are to win

Local pack: the map and short list of businesses Google shows for a nearby search. It has been the main local visibility prize for years, and it is far more generous than AI assistants.

SOCi’s 2026 Local Visibility Index analyzed roughly 350,000 locations belonging to 2,751 multi-location brands across 42 industry categories. On average, those locations appeared in Google’s local 3-pack 35.9% of the time. Gemini recommended 11.0% of them, and ChatGPT recommended 1.2%.

The implication is blunt. Ranking well in Maps does not guarantee a mention in an AI answer, and the AI list is much shorter. Note that this sample covers multi-location brands, so a single-site business may see different rates. If a competitor is named in the assistant’s answer and you are not, the customer may never see you at all.

What correlated with being recommended

Insites released a study on May 6, 2026, covering 10,000 US local businesses tested against ChatGPT and Perplexity. The company compared more than 200 digital marketing signals with real AI results. Its headline finding was that most businesses are known to AI but often described inaccurately: 93% were known, yet only 55% had accurate details.

The strongest links were ordinary local signals. Businesses found by both engines averaged 133 Google reviews, against 11 for businesses invisible to AI. Businesses in the local pack were recommended 3.5 times more often than those outside it. Insites also reported that a business’s own website was cited in 72% of awareness queries, and that AI-visible businesses had more than double the pages on average.

For a small business, the reading is practical. Reviews, a real local presence and a substantive website appear to matter more than technical tweaks. Insites reported that recently updated sites were more likely to appear, which is consistent with our look at content freshness and AI citations.

What each study measured

The two studies answer different questions, so they should not be merged into one statistic. The table shows the scope of each.

Study	Sample	Engines	Main finding	Caveat
SOCi 2026 Local Visibility Index	About 350,000 locations, 2,751 multi-location brands	Google 3-pack, ChatGPT, Gemini	3-pack 35.9%, Gemini 11.0%, ChatGPT 1.2%	Multi-location brands only; vendor-published
Insites AI visibility study (May 2026)	10,000 US local businesses	ChatGPT, Perplexity	133 reviews versus 11; local pack presence linked to 3.5 times more recommendations	Correlation, not causation; vendor sells audits

Both vendors sell products that depend on local AI visibility being a problem worth solving. That does not make the numbers wrong, but it is a reason to read the findings as hypotheses to test on your own business.

What a local business can do first

The evidence is early, so the sensible plan favors actions that help in classic local search too. None of these needs an AI-specific vendor.

1. Audit how ChatGPT, Gemini and Perplexity describe your business today, and write down any wrong hours, services or addresses. Insites found only 55% of known businesses had accurate details.
2. Fix the source of each error, starting with your Google Business Profile and your own website, since the website was cited in 72% of awareness queries.
3. Build a steady, honest flow of customer reviews. The 133-versus-11 gap is the clearest signal in the data, though it is correlational.
4. Add service and location pages with real detail, and keep them current.
5. Re-run the same prompts monthly, because answers vary, using the method in our guide to measuring AI search visibility.

Skip the shortcuts. Insites found structured data, FAQ schema and llms.txt showed weak or no correlation, which matches our earlier review of schema markup for AI search.

What the studies cannot tell you

Correlation is not a recipe. Businesses with many reviews are often larger, older and better known, and those traits may drive both reviews and AI mentions. Review volume may be a symptom of prominence, not the cause of a recommendation.

The engines also differ. Google’s AI features follow Google’s own rules, while ChatGPT and Perplexity draw on other sources, so a tactic that moves one may not move another. Our explainer on GEO, AEO and SEO covers why those surfaces differ. The honest position is to improve what is already good practice, measure it, and avoid paying for guarantees nobody can verify.

Frequently Asked Questions

Does a strong Google Maps ranking guarantee AI recommendations?

No. SOCi’s 2026 Local Visibility Index found the local 3-pack showed locations 35.9% of the time against 1.2% for ChatGPT, so the overlap is limited.

Are Google reviews worth collecting for AI visibility?

They are worth collecting for customers regardless. Insites found AI-visible businesses averaged 133 Google reviews against 11, though the link is correlational and not proven cause.

Should a local business add schema or an llms.txt file?

Not as a priority. Insites reported weak or no correlation between structured data, FAQ schema or llms.txt and AI recommendations in its 10,000-business sample.

Local AI visibility is still a young field, and most claims about it are stronger than the data. At Grow with Hafsa we help businesses build AI search and AI marketing programs that start with what the evidence supports, and we are glad to look at how assistants describe your business today.

What the GEO Research Actually Proves, and What It Does Not

Generative engine optimization (GEO) is sold with confident numbers. A July 2026 survey of the research behind it is more cautious. The strongest evidence applies to content an AI engine has already retrieved, not to whether it finds you in the first place.

Where the 40% figure comes from

GEO (generative engine optimization): the practice of increasing how often and how prominently your content appears in answers from generative search engines. The term was introduced by Pranjal Aggarwal and colleagues in a paper accepted to KDD 2024, a major data-mining conference.

The authors built GEO-bench, a collection of varied user queries across several domains, each paired with web sources able to answer them. Their abstract states that GEO methods can boost visibility in generative engine responses by up to 40%, and that effectiveness varies by domain.

Vendors often quote the 40% without the qualifier. The phrase is “up to,” the setting is a benchmark and the measure is visibility inside a generated response. That is a real result, but it is narrower than the way it is usually repeated.

What the 2026 survey concluded

On July 15, 2026, Olivier Martinez posted a critical survey to arXiv covering 45 studies published between November 2023 and July 2026. It frames GEO as a stochastic, partially observable pipeline: search activation, crawling, retrieval, reranking, citation and user behavior, each of which can change the outcome.

The central finding is a distinction. The original gains hold when a source is already in the engine’s context, so they do not show that a page will be found or that traffic will follow. The survey reports that no reviewed technique showed a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior, and it calls the overall evidence narrow.

This summary draws on the paper’s abstract and landing page, so per-technique effect sizes are not repeated here. Readers who rely on a specific number should consult the full text.

Which levers hold up

The survey does not say GEO is worthless. It says which findings travel and which do not. The table separates them in plain terms.

Finding in the survey	What it supports	What it does not support
Topical relevance and context position are the most reproducible levers	Writing tightly on-topic pages with the key answer early	Any claim that formatting alone creates citations
Generic heuristics transfer poorly across settings	Testing on your own queries and engines	Universal checklists that promise results
Competition can erode individual gains	Expecting returns to shrink as rivals adopt the same tactics	Permanent advantage from a one-time rewrite
Citation-oriented rewrites can impair retrieval	Caution before rewriting for style alone	Assuming every rewrite helps

The survey also reports high run-to-run variability in commercial audits, which is why our guide on measuring AI search visibility recommends repeated checks, not single screenshots.

What this means for your budget

The practical reading is to spend on what is both durable and testable. Clear, well-sourced pages that answer the question early serve classic search and AI engines alike, which is the argument made in our comparison of GEO, AEO and SEO.

Be more skeptical of tactics sold on a single mechanism. Our reviews of llms.txt and content freshness apply the same test: what was measured, on what sample, and does it show a causal effect.

When a vendor presents GEO results, ask four things.

1. Which engines and queries were tested, and how many repeated runs were used?
2. Were paraphrased queries and a control group included?
3. Does the result measure citation, or traffic and revenue?
4. Would the work still be justified if citation rates did not change?

Frequently Asked Questions

Does GEO work?

It works in the narrow sense that changes to already-retrieved content can change how it is cited. The 2026 survey found no technique with a proven lasting effect on being discovered or on traffic.

Is the 40% visibility gain a reliable benchmark?

It is the upper bound reported in the original paper’s benchmark tests. The authors also found effectiveness varied by domain, so it should not be read as an expected result for a given site.

What should a business do while the evidence is thin?

Invest in accurate, specific, well-structured content that helps readers, and measure citations and traffic over repeated runs. Avoid paying for guaranteed outcomes that no published study supports.

Thin evidence is a reason for disciplined testing, not for ignoring AI search. At Grow with Hafsa we help teams plan GEO and AI marketing programs around measurable pilots, and we are happy to look at which of your current claims the research actually backs.

Tracking AI Referral Traffic After ChatGPT’s UTM Tags Briefly Vanished

For about a week in October 2026, some ChatGPT visitors stopped looking like ChatGPT visitors in analytics. The episode shows that AI referral traffic is a floor, not a full count, and that any report built on a single tag can break without warning.

What was reported in October 2026

On October 14, 2026, Barry Schwartz of Search Engine Roundtable reported that ChatGPT links appeared to have lost their UTM parameters: short tags added to a URL, such as utm_source=chatgpt.com, that tell analytics where a click came from. The change seemed to coincide with the GPT-6 release about a week earlier.

Cedric Huesler of Adobe analyzed about 500 pages that normally received at least 5% of their visits from ChatGPT. Comparing October 9 to 13 with the prior week, he found ChatGPT visits down by about 31,000 a day and Direct visits up by about 26,000 a day. Search traffic on those pages fell 7%, in line with other pages. Direct rose on two-thirds of the pages that lost ChatGPT traffic.

The reading is straightforward: the visitors did not disappear, they lost their label. On October 15, Huesler said it was a bug and that tagging had resumed. The report contains no OpenAI statement, so the cause is practitioner-reported, not confirmed by the vendor.

How AI platforms tag their links

Attribution depends on two signals: a UTM tag in the URL and a referrer, the header a browser sends to say which page the click came from. Seer Interactive tested this across client properties, in an article originally published in December 2024 and updated June 22, 2026, and did not publish a site count.

Its findings differ by platform, which is why a single rule rarely covers everything. Seer also notes that privacy-sensitive queries and account-specific pages usually omit UTMs, and that tags are more likely on publisher citations than on other links.

Source	UTM tag present?	Referrer present?	Seer Interactive finding
ChatGPT citation and source links	Usually, as utm_source=chatgpt.com	Varies	Most likely place to find the tag
ChatGPT search results	Usually not	Usually yes	Attributed through the referrer
Gemini, Perplexity, Claude	Not in Seer’s testing	Reported as present	Rely on the referrer alone

The practical lesson is that two separate mechanisms can fail independently. When both fail at once, as reported in October, the click is filed under Direct.

What GA4 does natively

Google’s Analytics Help documentation now defines an AI Assistant channel in the default channel group: “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” The same page excludes Google’s AI Overviews and AI Mode from that channel, so those clicks are not separated from regular Google organic traffic here.

Google does not state a start date on that page, and it does not publish a full list of recognized sources. The channel is assigned from the referrer, so it cannot rescue a click that arrives with no referrer and no tag.

For a business reporting to a board or a client, this means the native channel is a useful default but not a complete measure. A custom channel group, built on the source domains you care about and placed above Referral in the rule order, gives you control over what counts.

A defensible way to measure it

The goal is not a perfect number. It is a number you can explain and compare week to week. The following sequence works for most small teams.

1. Create a custom channel group in GA4 that matches the source domains of the AI platforms you care about, and keep it in addition to the native AI Assistant channel.
2. Record the date you set it up, so later changes in the data can be separated from changes in tagging.
3. Chart Direct and AI traffic together for your top landing pages. A fall in one with a matching rise in the other is the October pattern.
4. Report AI traffic as a floor, with a note that untagged visits sit inside Direct.
5. Pair the traffic view with citation checks, using our guide to measuring AI search visibility over repeated runs.

Conversion quality matters as much as volume. Our review of three studies on AI search conversion explains why a small, well-labeled segment can still be worth tracking carefully.

What to ask whoever reports your AI traffic

If an agency, a tool or an internal team reports AI traffic to you, ask how the number is built. A credible answer names the source domains matched, states whether Direct is reviewed, and says what happens when a platform changes its tagging.

A report that shows AI traffic jumping or collapsing in a single week deserves a question about tagging before it deserves a strategy change. The October episode produced exactly that pattern without any real change in visitor behavior.

The same discipline applies to agentic reporting, where software compiles the numbers on your behalf. An agent can only be as accurate as the labels in the data it reads, so the review step stays human.

Frequently Asked Questions

Did ChatGPT stop sending traffic in October 2026?

The reported evidence points to a labeling problem, not a traffic loss. Huesler’s analysis found Direct rising by roughly the amount ChatGPT fell, and he later described the cause as a bug.

Does GA4 track ChatGPT traffic automatically?

Partly. GA4’s default channel group includes an AI Assistant channel that covers sources like ChatGPT, Gemini and Copilot. It depends on the referrer, so visits without one still appear as Direct.

Is AI referral traffic worth tracking if the volume is small?

Yes, as a separate segment. It is a leading indicator of whether your content is being cited, and its conversion behavior may differ from search traffic. Track it as a range and compare it over time.

Reliable AI traffic reporting is a measurement design problem before it is a marketing one. At Grow with Hafsa we help teams set up AI search measurement and AI marketing reporting that holds up when platforms change their tagging, and we are glad to review your current GA4 setup.

GEO, AEO and SEO: What Each One Optimizes For and Where They Overlap

Three acronyms now compete for the same budget line: SEO, AEO and GEO. They overlap heavily, but each targets a different surface and is judged by a different result, and knowing which is which keeps you from paying twice for the same work.

What SEO still covers

SEO (search engine optimization): the work of making pages crawlable, relevant and trustworthy enough to rank in a search engine’s results. Its success measures are rankings, clicks and the traffic and enquiries that follow.

SEO also remains the entry ticket for Google’s AI features. Google’s documentation on AI features states that to be shown as a supporting link, “a page must be indexed and eligible to be shown in Google Search with a snippet,” and that there are no additional technical requirements.

The implication is practical. If your pages are not indexed, no amount of AI-specific tactics will place them in Google’s AI answers. Fix crawling, indexing and snippet eligibility first.

What AEO adds

AEO (answer engine optimization): shaping content so a system can lift a short, direct answer from it. The older form is the featured snippet, and the newer form includes answers spoken or summarized by AI assistants.

Google’s featured snippets documentation describes snippets as special boxes where the format of a regular result is reversed, and says site owners cannot mark a page as a featured snippet. Google’s systems decide whether a page makes a good one for a given query.

So AEO is less a technique you switch on than a writing habit. Put the direct answer first, define terms in one sentence, and make each section understandable when quoted alone. Our article on query fan-out shows why answering the related sub-questions matters as well.

What GEO research found

GEO (generative engine optimization): the practice of increasing how often and how prominently your content appears in responses from generative search engines. The term comes from a paper by Pranjal Aggarwal and colleagues, accepted to KDD 2024.

The authors built GEO-bench, a benchmark of 10,000 queries, and tested nine content methods. Their abstract states that GEO “can boost visibility by up to 40% in generative engine responses.” The strongest methods were adding citations to sources, adding quotations and adding statistics, while keyword stuffing performed poorly and measured about 10% below baseline on Perplexity.

The paper also reported that results varied by domain and that lower-ranked sites gained more than top-ranked ones. These are results from the authors’ own tests, not guarantees for live engines. They do support a simple habit: make claims specific, sourced and quotable.

The three disciplines side by side

The table summarizes how the disciplines differ in target, measure and levers. The final column notes how strong the public evidence is for each.

Discipline	Target surface	Typical success measure	Main levers	Evidence base
SEO	Ranked results pages	Rankings, clicks, conversions	Crawlability, relevance, links, page quality	Mature, documented by search engines
AEO	One extracted answer, such as a featured snippet	Snippet or answer appearances	Direct answers, clear definitions, structure	Documented for snippets; engine-selected
GEO	Sources cited inside AI responses	Citation frequency and position	Sourced claims, quotations, statistics, clarity	Early; one peer-reviewed benchmark plus vendor studies

Notice that Google’s AI features page does not use the terms GEO or AEO at all. The labels come from the industry, which is why they are defined differently from vendor to vendor.

Where to start, in order

Because the disciplines share most of their foundations, sequence matters more than choosing a label. A workable order for a small team follows.

1. Confirm important pages are indexed, snippet-eligible and not blocked from the crawlers you want, using our guide to robots.txt for AI crawlers.
2. Rewrite key sections so the direct answer comes first and each passage stands alone.
3. Add attributed statistics and named sources inline, following the pattern the GEO paper tested.
4. Treat schema as supportive only, as outlined in what the evidence supports for schema markup.
5. Measure citations over repeated runs, since answers vary, as described in our guide to measuring AI search visibility.

What to ask a vendor who sells GEO

A vendor selling GEO as a separate product should be able to explain what it does that good SEO does not. Google states there are no special optimizations necessary for its AI features, so any claim of a secret technique needs evidence.

Ask which engine the work targets, how citations are measured across repeated runs, and what happens to results when the engine changes. Ask for a before-and-after example with the method stated. If the answer is a proprietary score with no method, treat it as marketing, not measurement.

Our piece on how to choose an AI marketing agency adds general questions that apply here too.

Frequently Asked Questions

Is GEO different from SEO?

GEO targets citations inside AI-generated answers, while SEO targets ranked results. They share foundations such as indexing and content quality, so most of the work overlaps.

Does Google require special optimization for AI Overviews?

No. Google’s documentation says there are no additional requirements or special optimizations needed beyond a page being indexed and eligible to appear with a snippet.

Which discipline should a small business prioritize first?

Start with SEO fundamentals, because indexing and quality gate everything else. Then structure content for direct answers and add sourced evidence, which serves AEO and GEO together.

Separating these labels makes budgets easier to defend and vendor claims easier to test. At Grow with Hafsa we help businesses plan AI search and AI marketing work around evidence instead of acronyms, and we are happy to review where your current effort overlaps.

For a closer look at how far the research on GEO actually reaches, see what the GEO research actually proves, and what it does not.

Google’s Business Agent for Leads: What a Chat Inside the Ad Means for Lead Generation

On May 20, 2026, Google announced a Business Agent for leads: a chat agent inside a search ad that answers questions from the advertiser’s own website instead of sending people to a static form. For lead-generation advertisers, the first sales conversation moves into the ad itself, before anyone reaches your site.

What Google announced

Google’s Ads blog post, “A new generation of ads for the AI era of Search,” published May 20, 2026, says the Business Agent for leads “puts a smart brand agent right inside your ad.” Built with Gemini, it lets a person click “Chat” and get answers rather than fill in a static form.

Google’s example is a student researching universities who receives instant answers based on the university’s website. The post says the format is coming “in the coming months,” so its availability in your account may differ from what this article describes.

The shift matters because a form captures contact details and little else. A conversation can answer objections, narrow intent and pass a warmer lead to your team, if the agent is set up well.

Your website becomes the knowledge base

The detail that deserves the most attention is the source of the answers. Google states that the agent’s answers are based on the advertiser’s website. Whatever your pages say, or fail to say, is what a prospect will hear in the chat.

This is a claim-then-consequence situation. A page that hides pricing ranges, leaves out who the service is not for, or contradicts another page gives the agent nothing reliable to repeat. Pages that state scope, eligibility, timelines and exclusions plainly give it material it can quote accurately.

For service businesses, that means the audit starts on the site, not in the ad account. Our guide to AI crawler access covers the related question of which automated systems can read your pages at all.

Static form versus chat agent

The table below compares the two formats. The first column reflects Google’s description; the remaining rows are our reading of what each format implies for a lead-generation team.

Question	Static lead form	Business Agent chat
Where do answers come from?	No answers; the form only collects details	The advertiser’s website, per Google
What does the prospect do?	Fills fields and waits	Asks questions and gets replies in the ad
What can go wrong?	Low-intent submissions and slow follow-up	Inaccurate or off-brand answers drawn from unclear pages
What your team must own	Follow-up speed	Page accuracy, qualification rules and handoff

The risk moves from the form to the content. A weak form wastes a lead; a weak knowledge base can misinform one.

What Direct Offers shows about Google’s direction

Direct Offers: a Google pilot, begun in January 2026 with brands including Chewy, Gap and L’Oreal, in which advertisers upload promotions and Gemini builds a deal for the shopper. In the May 20 post, Google says brands supply the promotions, the eligible products and guardrails, and that “we use Gemini to construct a deal.”

The pattern is the same as with the Business Agent. The advertiser provides inputs and limits, and Gemini composes the customer-facing exchange. Google also says that to take full advantage of the new formats, advertisers should build a strong foundation with AI Max for Search, AI Max for Shopping and Performance Max.

Google cites a December 2025 Ipsos Global Consumer Journeys survey it commissioned, covering online shoppers aged 18 and over in 25 countries who use AI Overviews or AI Mode for shopping. In it, 75% said they make faster, more confident decisions. The sample is shoppers, not B2B buyers, so treat it as directional. For background on the campaign type involved, see our article on Google AI Max for Search.

How to prepare before it reaches your account

Because Google has not described advertiser controls, preparation should focus on what you can govern yourself. Work through these steps in order.

1. Audit the pages an agent would draw on, and fix contradictions in pricing, service area, timelines and eligibility.
2. Write down what a qualified lead looks like in plain language: budget range, location, company size and urgency.
3. Decide the handoff: who receives a chat that turns serious, how fast, and through which channel.
4. Choose a lead-quality measure beyond form fills, such as qualified meetings, and connect it to your CRM.
5. Start with a capped test budget and review chat transcripts weekly, if the account exposes them.

The same discipline applies to any agent that speaks or spends for you. Our piece on approval gates for agentic ad spend explains how to decide where human sign-off belongs.

What remains unconfirmed

Several practical questions are unanswered in the announcement. Google does not state which industries will receive the format first, whether advertisers can edit or restrict what the agent says, how transcripts are shared, or how it is billed.

That gap is the reason to treat vendor pitches carefully. Anyone claiming to know the controls, the conversion lift or the pricing before Google has published them is guessing. Our overview of what is actually autonomous in agentic advertising offers a way to test such claims.

Frequently Asked Questions

What is Google’s Business Agent for leads?

It is a Gemini-built chat agent inside a search ad. Instead of filling in a form, a person can click “Chat” and receive answers based on the advertiser’s website, according to Google’s May 20, 2026 announcement.

Can advertisers control what the agent says?

Google has not described advertiser controls in the announcement. Until it does, the safest lever is the accuracy and clarity of your own website pages.

Is the format available now?

Google said it was coming “in the coming months” as of May 20, 2026. Check your Google Ads account or your account representative for current availability.

Preparing a site and a qualification process for agent-led conversations is practical groundwork, and it improves ordinary lead handling too. At Grow with Hafsa we help teams build AI marketing and performance-marketing workflows that keep a person in charge of the decisions that matter, and we are glad to talk through your setup.