Can a business earn a mention in Google AI Overviews? A reproducible 90-day test

Yes, businesses do get named in AI Overviews, and no publisher can promise you will be one of them. The only honest way to talk about it is a rate with both its numbers visible: businesses named in the answer text, divided by eligible business-query observations, with the study dates attached. This page publishes that study design in full, including the codebook, so anybody can run it. As of 29 September 2026 we have not completed a 90-day AI Overviews study, so there is no named-business rate on this page, and the results section says so rather than filling the gap.

What we can give you, dated, is what happened on three other assistants in the same kind of work, including the one publish-to-citation event in our files that a reader outside the company can check.

What "recommended", "named", and "cited" mean

These three words are used interchangeably in most writing about AI search, and the measurements they imply are not the same. Keep them apart or your numbers will mean nothing.

• Named. The business name appears in the answer text itself. This is the primary measure, because it is what a reader sees without clicking anything.
• Cited. A link to the business's domain appears among the sources shown with the answer. A business can be cited without being named in the text, and named without being cited.
• Recommended. The answer presents the business as a suggested option, rather than mentioning it as an example, as a competitor, or as something to be careful about. This needs a human judgement and therefore needs a codebook.
• Eligible business-query observation. One run of one query, in a category where a company-level answer is plausible, where an AI Overview actually appeared. If no overview appeared, the observation is logged and excluded from the denominator, and the exclusion count is published.

The headline figure is businesses named in answer text divided by eligible business-query observations. Named and cited and recommended are each reported separately underneath it.

What Google officially says about eligibility

Google publishes guidance for site owners about AI experiences in Search, and it is worth reading directly rather than through summaries. The parts that matter for a business are the following four, and none of them amounts to a formula.

• Ordinary indexing applies. Content has to be crawlable and indexable before it can be used at all, which is a lower bar than most businesses verify they clear.
• Snippet and preview controls exist, and using them restrictively can keep you out of the answer surface as well as out of a snippet.
• The usual content guidance applies: content made for people, with something specific in it, rather than content assembled to appear in a search feature.
• No eligibility criterion for business recommendations is published. We are not aware of a Google document stating what makes a company get named, and we treat lists of "AI Overview ranking factors" as inference, including anything we infer ourselves.

Read Google Search Central's documentation on AI features and its general site owner guidance, and write the date next to anything you quote. These pages are revised.

Study design

The nine fixed choices

• Query set: 120 queries, frozen before day one, in six business categories of 20 each, so that no single sector drives the headline.
• Query types within each category: selection queries, constrained selection queries with a city or budget, task queries, comparison queries, price queries, and risk queries. Record the type per query, because the rate differs enormously by type and a blended figure hides it.
• Schedule: every query run once a week for 13 weeks, giving up to 1,560 observations.
• Location and language recorded for every run, because both change the answer.
• Signed-out sessions, to reduce personalisation, with the browsing state recorded.
• Capture the whole answer, the full source list, and a note of whether an overview appeared at all.
• Two coders, with agreement checked on a sample, because "recommended" is a judgement.
• Exclusions published, not quietly dropped: queries where no overview appeared, where the answer was a refusal, or where the page failed to load.
• Nothing named. No brand appears in any query, including your own.

The codebook, which is the part that makes it reproducible

Six rules, written down before scoring starts and not changed mid-study.

• A business name counts as named only if it identifies a specific company, not a category or a generic phrase.
• A government body or official portal is coded as an official source, not as a business, and counted in its own column.
• A business named only inside a source link title, and not in the answer text, is cited and not named.
• "Recommended" requires the answer to present the business as an option for the asker. A mention as a cautionary example is coded as named and not recommended.
• Multiple businesses in one answer are all counted, and the position of each is recorded.
• Any coding decision that takes more than thirty seconds gets written into the codebook as a new rule, with the date, so the same case is decided the same way later.

Results by page and business type

We have not run this study. There is no named-business rate, no breakdown by query type and no breakdown by business type on this page, as of 29 September 2026. When it is run, this section will carry the numerator, the denominator, the exclusions, and a separate count for each of the six query types and each of the six categories, with the run dates.

What we have completed, all on other assistants, with dates, is this.

• ChatGPT, 27 July 2026, on clawlaw.in. 78 blind buyer questions, and the assistant confirmed afterwards that it had run no live web search for any of them. Breadth of naming in that memory-based reading was ProVakil on 28 questions, CLAW on 21, Legistify on 18.
• ChatGPT, 6 August 2026, on clawlaw.in. clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india cited as a source for a vendor due diligence question, fourteen days after that page was published, in a zone where the same question set had named the company nowhere at baseline. This is the only result in our files with a URL a stranger can check.
• ChatGPT, 18 August 2026, on clawlaw.in. Across eighteen blind commercial questions the company was the top source on exactly one of them, and its own comparison page was named in the answer as still being the vendor's own editorial page.
• Claude, 17 September 2026, on clawlaw.in. Across six blind commercial questions, official court portals took every position above any commercial product on the how-to questions, no third party review or directory source made it into any answer, and two enterprise vendors outranked the company because they publish explicit court and tribunal coverage lists.
• Perplexity, September 2026, on aiknowsus.com. Across 24 batches and 72 conversations on our own domain, the phrase recording that we were not cited appears 161 times in the engines' own self audits. In one batch of six questions the assistant reported it had not cited or recommended aiknowsus.com in any of the six answers.

What changed during the 90 days

This section exists in the design because a 90-day window always contains changes that are not yours, and a study that does not log them will attribute them to itself. Record these six, with dates, as they happen.

• Product changes on the answer surface, including which queries show an overview at all.
• Model changes announced by the provider.
• Your own publishing: every page published or edited, with its URL and date.
• Competitors' publishing, where you notice it.
• Changes on official or authoritative sources in your category, which can displace commercial results entirely.
• Anything that broke on your own site, including a rendering change that stops facts reaching a crawler.

We have a dated reason for the last one. A pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of from the company's own site. Claude, 6 August 2026, on clawlaw.in. A change like that inside a study window would look like a ranking change and would not be one.

Practical checklist for business owners

Eight things, each of which is defensible from something in our own logs rather than from a theory.

• Check a machine can read your commercial facts with scripts disabled. This has already failed once in our programme.
• Publish a price or a band with a date. Undated and absent prices both get filled in from somewhere else.
• Make your coverage a named, counted list. Coverage lists decided an ordering against us on 17 September 2026.
• Source and date every figure about a competitor. Correct figures without a source were treated as advocacy.
• Keep one version of each fact in public. Two price lists in public were noticed and called out on 6 August 2026.
• Delete any claim that could be shown to be impossible. One cost a company the answer entirely.
• Write the limits page. Who should not buy from you, and why.
• Never address the model in hidden text. It was found, refused, and the company was named for it.

None of that buys a mention. It removes the reasons we have actually recorded for not getting one, which is a different and more honest offer.

What this study cannot show

• It cannot show cause. Ninety days contains model changes, product changes and everybody else's publishing.
• It cannot generalise past its query set. A different 120 queries is a different study with a different rate.
• It cannot escape location and personalisation completely, even signed out.
• It cannot compare with other published rates unless those publish their queries, codebook and exclusions.
• It cannot promise anybody a mention, and a supplier who does is selling something they do not control.
• It cannot tell you about ChatGPT, Claude, Gemini or Copilot, each of which needs its own run.

Common questions

Why publish a study design with no results in it?

Because the design is the part that is reusable, and because a rate without a method behind it is not evidence. We would rather be the page that lets you run this than the page that tells you a number you cannot check. When we have the numbers they will go in the results section with their dates.

Are AI Overviews the same thing as an assistant answer?

No. They are a search surface with their own selection behaviour, and results from an assistant do not carry across. Every measurement on this site is reported per engine for that reason.

Does blocking AI answers protect my content?

Snippet and preview controls exist and using them restrictively can remove you from the answer surface as well. That is a business decision with a real cost either way, and it should be made deliberately rather than by copying somebody's configuration.

How many queries do I need if I am a single local business?

Twenty to thirty of your real buyer queries, each with your city in half of them, run weekly for a month. That is small, it will not produce a publishable rate, and it will tell you whether businesses like yours get named in your category at all, which is the thing you need first.

What is the single most common reason a business is absent?

On our logs, not being findable as a set of checkable facts. Prices a crawler cannot see, coverage stated as "comprehensive", comparisons with no sources, and contradictions between a website and a listing. Every one of those is dated in our capture files and every one of them was the company's own to fix.

Can you guarantee we will appear?

No. Nobody can, and we will not write a sentence that implies it.

Dataset, codebook, and corrections

To run this you need four artefacts and no software: a frozen query sheet with 120 rows and their types, an observation log with one row per run carrying location, language, date, time zone, whether an overview appeared, the businesses named, the sources shown and the coding decisions, a codebook file with the six rules above plus every rule added later with its date, and a folder of saved answers.

Our own observations come from the clawlaw.in programme from July 2026, with dated readings on 27 July, 6 August, 18 August and 17 September 2026, and from the aiknowsus.com audit of September 2026 covering 24 batches and 72 conversations. Of these, the observations concerning public pages and the 6 August citation can be checked from outside. Counts taken across our capture files cannot be until the captures are published, and we say so on every page that uses them.

Corrections and change log. First published 29 September 2026, with the design and codebook complete and the results section empty and marked empty. Corrections to this page will be listed here with their dates, and the original wording will stay visible rather than being edited away. If you find an error, or a Google document that does publish eligibility criteria for business mentions, we want to hear about it.

Disclosure. Written by AI Knows Us, which sells AI visibility measurement, so read the recommendation to measure as coming from an interested party.

What to do first

Pick ten queries your buyers actually type, run them signed out this week, and for each one record only three things: did an overview appear, was any business named, and was it you. Ten rows will not be a study. It will tell you within an hour whether your category produces company-level answers at all, and that decides whether the 90-day version is worth your time.