Thirty Local Service Questions, and Google Showed No AI Overview Once

We asked eight versions of “which local service business is best” across Spain, the United States and Germany. Across sixteen runs, Google’s AI Overview did not appear once. AI Mode answered all sixteen, on the same call, in the same instant, with a full set of citations. Four control questions, local in exactly the same way but asking for information rather than a recommendation, returned an AI Overview on all eight of their runs. The question is not who gets cited. For this shape of question, on that surface, there is nothing to be cited in. Take the same question into education or hospitality, though, and an overview usually does appear: that boundary is measured below.

That distinction is the whole piece. A visibility report that says a local business is cited zero times in AI Overviews is telling the truth and telling you nothing, because no local business is cited there. The surface does not run.

Disclosure: EchoWi sells AI visibility measurement, so a study arguing that some AI visibility numbers describe a surface that never rendered is a study with an interest, and it also cuts against the simplest version of what we sell. Everything needed to repeat it is here: the eight questions and the four controls are printed in full, the surfaces are Google’s AI Overview and AI Mode, the markets are Spain, the United States and Germany in their own languages, each was run twice, the date is 11 August 2026, and every row is in our measurement register.

The short version

1. Eight local buying questions, three markets, three languages, zero AI Overviews. Not a low number. Zero runs out of sixteen produced one.
2. AI Mode answered all sixteen. Same call, same moment, so this is not the tool failing or the market being quiet.
3. Four controls rule out the obvious objection. Questions just as local but informational, one in each of the three markets and a second in Spain, returned an AI Overview on all eight runs.
4. Dropping the word “best” changes nothing. The same intent asked as “which one do you recommend” also returned no overview, so this is not a superlative being filtered and cannot be rephrased around.
5. An independent path confirms it. The live search results page for one of the commercial questions contains six local pack blocks and no AI Overview element at all. The control’s results page carries an AI Overview in first position.
6. So there are two different zeros and most tools report one number. “The engine answered and did not cite you” and “the engine did not answer” are different facts with different responses, and only one of them is about your content.

What was asked, and what came back

Each question went to AI Overview and AI Mode in the same call, twice, in its own market and language.

One run of the Barcelona question errored rather than returning an empty answer. That is recorded separately in the register, because an error is the instrument failing and an empty answer is the surface declining, and the difference between those two is the subject of this study.

The first question was also asked a third time with three runs requested. AI Overview answered none of them.

Why the control is the load-bearing part

A scraper that quietly fails on one shape of query would produce exactly the table above, and we have published that mistake before in a different form, so the zero was not taken at face value.

The controls are questions that are local in exactly the same way, name a city, were asked in that city’s language and market on the same day through the same tool, and differ only in asking for information rather than for a recommendation. All four returned an AI Overview on both of their runs, eight for eight. There is one in every market the commercial questions were asked in, so the demonstration that the instrument renders overviews for local questions is not confined to any of them.

So whatever is suppressing the overview is not Spanish, not place names, not the market, and not the instrument. It tracks the thing the question asks for. And it is not about regulated professions. The first six questions are all health, legal or financial services, which invites the reading that Google is being careful with advice rather than with local recommendation. So two more were asked in verticals nobody regulates: the best Italian restaurant in Madrid and the best plumber in Seville. Neither returned an overview either, and AI Mode answered both, citing Michelin and two food blogs for one and three trade marketplaces for the other.

Then a second, independent path. Rather than asking a surface, we fetched the live search results page and looked at which blocks Google actually placed on it.

That is the mechanism in one table. Google is not withholding an answer to the local buying question. It is answering it with a different product: the map, the three business cards, the review stars. The generative answer is what it uses when the question wants prose.

A wider sweep, and an idea that did not survive it

The study above asks a surface whether it answered. The results page can be asked directly whether Google placed an AI Overview block on it at all, which is the more direct evidence of the two, and the two agree where they overlap.

So 22 more local buying questions went out that way: 8 Spanish verticals in eight cities, from hairdressers to vets to estate agents, 9 American ones and 5 German. None of the 22 returned an AI Overview. All 22 returned a local pack, between 3 and 7 business blocks each.

Counting both paths together, that is thirty distinct local service questions across three markets and seventeen verticals, and not one AI Overview. Ten control questions were asked alongside them and nine returned one.

The tenth is the interesting part, and it killed a better-sounding version of this finding. Asked for the opening hours of a municipal library in Zaragoza, Google returned a local pack and no overview. That suggested a sharper rule than the one published here: not “the question asks which business to choose” but “the question resolves to a place”. It is a tidier idea and it would have covered both results.

The American equivalent, opening hours of the Denver public library, returned an AI Overview and three local packs.

One market’s worth of evidence was enough to state that rule and not enough to keep it, so it is recorded here as dead rather than quietly dropped. What survives is the narrower claim, which thirty questions have now failed to break: ask which local business is best, and on that surface there is nothing to be cited in.

Where this stops being true, which we went looking for

Everything above is trades and services: plumbers, dentists, lawyers, gyms, accountants, vets, hairdressers, estate agents, restaurants. That is a wide sample and it is not the whole of local commerce, so the next thing to do with a finding this clean is to take it somewhere it might break.

It breaks in education and hospitality.

Six of eight, and ten more questions in the same two verticals then split five and five, which puts education and travel together at eleven overviews in eighteen questions. That is not a side of a boundary, it is a coin flip.

“Best hotel in Chicago” and “best private university in Boston” are “which one in this city is best” in exactly the form that returns nothing for a plumber, and both return an overview, one of them alongside three local packs.

So the honest shape of the finding is narrower than the headline above and worth stating plainly. Across thirty questions in trades and professional services, in three markets, no AI Overview appeared. In education and hospitality, most of the time one does. The suppression is a property of a class of categories rather than of local buying questions in general.

We do not know what separates the two classes, and the tempting explanation does not survive contact either. The obvious guess is that categories with an established ranking industry, university league tables and hotel guides, give the model something to summarise. But restaurants have Michelin and return nothing, and a hotel in Seville returns nothing while a hotel in Chicago returns an overview. Whatever the rule is, it is not that, and one day of data is not going to find it.

The second tidy explanation died the same way as the first, and it is worth naming because it is the one the table above invites. Within the ten new questions, every one naming a city returned nothing and almost every one naming a country returned an overview, which looks like a rule. Checked against all eighteen it is not one: a university in Madrid, a university in Boston and a hotel in Chicago each name a city and each return an overview.

The operating consequence is unchanged and, if anything, sharper. You cannot infer this for your own category from anyone else’s. Ours is the only sample we have, it splits down the middle between two groups of verticals, and the check that settles it for you is one query.

The comparison the study was missing, and the mechanism it found

Every services question above names a city. That means “no overview for services” and “no overview for local questions” were the same rows, and nothing here could tell them apart. So the same trades were asked without a place: best online therapy service, best online bookkeeping service, best online legal service for a will, best virtual assistant service, and two Spanish equivalents.

All six returned an AI Overview. The same professions that return nothing when a city is named. Take the place out and the answer appears.

That points at something checkable, so it was checked against every commercial question in this study rather than the batch that suggested it. Counting whether Google placed a local pack against whether it placed an overview:

Forty-six questions, and in forty-one of them exactly one of the two appears. The local pack and the AI Overview behave like alternatives, and the suppression everyone would describe as “Google does not answer local buying questions” is better described as Google answering them with a different module.

The exceptions are named rather than rounded away. Two questions carry both, a university in Boston and an all-inclusive resort in Mexico. Three carry neither, including a hotel in Seville. Five out of forty-six is not a rule with no exceptions, and this is the first pattern in this study to survive being tested against everything measured before it, after four that did not.

If it holds, the practical version is short. Look at your own query and see which module Google places. A local pack means the citation question is the wrong question and your Google Business Profile is the surface. No local pack means there is an answer to be cited in, and it is worth measuring who is in it.

The part that changes what you should buy

If you run a dental clinic, a law firm or an accountancy practice, the single most valuable question in your category is “which is the best one near me”. Everything in AI visibility marketing is aimed at making you the answer to it.

On AI Overview, in three markets, that question has no answer to be part of.

This does not mean local businesses are invisible in AI search. It means the surface where they are visible is a different one, and it behaves differently. AI Mode answered all eight questions and cited between four and fourteen domains each time. That is a real, reachable, winnable surface. It is simply not the one most reporting is built on, because AI Overview is the one that is easy to scrape at volume.

Two consequences follow, and the second is the expensive one.

A zero needs a denominator before it means anything. If a report tells you that you appear in zero AI Overviews for your local buying question, ask how many AI Overviews it saw for that question. If the answer is zero, the finding is about Google’s layout and not about you, and no amount of content will move it.

The competitor set on the surface that does answer is not the one you expect. In the Austin and Munich dental questions, the domain cited in every AI Mode run was www.google.com itself. In Valencia the durable citations were a clinic’s own site and a directory. In Madrid, one law firm’s own domain held the only stable slot across both runs. Those are three different competitive situations in one vertical, and none of them is visible from a surface that never rendered.

This is the same problem in a different costume to a crawler that never reaches your pages, where a blocked bot and an unpersuasive page produce the same silence. It also sits underneath the finding that fewer than three in ten durable citations belong to the company selling the product: before asking who holds the slot, it is worth checking that the slot exists.

The most reliable citation in a local answer is Google

AI Mode answers these questions, so the next question is who it cites, and one domain turns up more than any business does.

Reading ten local buying questions this way, www.google.com is present in every run of six of them. In the Austin and Munich dental questions it is the only domain that is. It is not a coincidence of one vertical: it appears for a dentist, a restaurant, a hairdresser and a vet, across three markets.

What is being cited there is Google’s own place data, the Maps entry and the business profile. Which makes the practical consequence unusually blunt for a finding this early: for a local business, the most reliably cited source in the answer about your category is not anybody’s website. It is the profile you fill in on Google, and the reviews sitting under it.

That reframes what “AI visibility” means for a plumber or a clinic. The extractable answer block, the schema, the FAQ, all the work that pays in the software categories we measured, is aimed at a slot that in six of these ten answers is occupied by Google quoting itself.

There was a tempting pattern in the four it was missing from: the six with Google were consumer, walk-in businesses and the four without were professional services, a law firm, two accountancies and a plumber. A clean story on ten data points, which is precisely why it got tested rather than published.

It is dead. Two more questions, one from each side of the supposed line: a gym in Valencia carries Google, as predicted, and an architecture studio in Seville, a professional service that should not, carries it in both runs. Twelve questions now, eight with Google, and no rule.

That is the third tidy explanation this study has killed in a day, after “the question resolves to a place” and “naming a city suppresses the overview”. Each one fitted the batch that suggested it perfectly. Each one failed against everything measured before it. At some point the pattern of failures is itself the result: the properties a marketer would reach for to predict any of this, the vertical, the place, the kind of business, do not predict it. The only thing that has told us anything about a category is running the query for that category.

What AI Mode cited instead

Worth reading on its own, because it is where the answer to these questions actually lives.

Valencia, dental. Eight domains, three of them in both runs: a clinic’s own site, a clinic directory, and Google. Doctoralia and TopDoctors appeared once each. This is a mixed field where an individual practice can and does hold a stable citation.

Madrid, employment law. Thirteen domains, one of them in both runs: a single firm’s own site. The rest, including a ranking site, a news outlet and eight other firms, each appeared once. One firm has a durable position and everyone else is rotating.

Austin and Munich, dental. Four and six domains, and the only one present in both runs was Google’s own. Beyond that, a review platform, a consumer association, a listings site and individual practices, each appearing once. Nothing local held a stable position in either city.

The pattern across the eight is that the durable slot, where there is one, went to a single practice or to a directory, never to more than one competitor at a time. That is a much more concentrated situation than the software categories we measured last week, where the stable core ran to seven, thirteen and twenty-three domains.

The rate held and the exception moved

This study’s one surviving pattern is that the local pack and the AI Overview behave like alternatives, and the limits below say plainly that it describes one day’s layout. Five days later is the cheapest test of that sentence, so eight of these questions went back to the results page, four from each side. Seven answered and one returned a server error, recorded as an error rather than as an absence, because the difference between those two is what this whole piece is about.

6 of the 7 carry exactly one of the two blocks. The rate holds.

2 of these questions can be compared strictly, and 0 of the 2 stayed on the same side. Only best private university in Boston and best hotel in Chicago have a recorded results-page row from 11 August; the other five were measured that day by asking the AI surfaces, not by reading the page, so today’s reading of them is fresh evidence about the layout rather than a repeat of a recorded one. Of those two:

Boston was one of 3 rows in the 52 read that day carrying both blocks at once, which is what made it an exception. It is not one now. Chicago was not an exception and is one now. The exception did not go away, it moved to a different question, and the headline rate stayed where it was through both moves.

That is two cells and it is published as two cells. What it is enough to say is that the exception is not a property of the question. A page that carries both blocks today is not a page that reliably carries both, and a page that alternates today was not necessarily alternating last week. If you are watching one query to decide whether a local answer has room to be cited in, you are watching something that changed for both of the two queries we can check.

This register already carries the same shape from a different study: a share can hold steady while the thing underneath it turns over completely. There the share was citation stability and the turnover was the set of cited domains. Here the share is the alternation and the turnover is which question is the exception. Two dates are not a trend and this is not offered as one.

Limits, and one of them is large

Two runs. Every row here is two runs per surface, and one question got three. For the AI Overview result that is enough, because the finding is a complete absence rather than a rate: sixteen runs across eight questions produced no overview at all. For the AI Mode citation sets it is not enough, and those are reported as observations rather than rates.

Eight questions. Local commercial intent is a huge space and this is eight points in it, chosen across five verticals and three countries to avoid a single-market artefact. It is not a survey of local search.

2 dates, 5 days apart. Google’s layout decisions change. The sweep describes 11 August 2026 and 7 of its questions were read again on 16 August, which is what the section above reports. Two dates are not a trend, and the same questions could return overviews next month.

Four controls, three markets. They establish that the instrument renders overviews for local questions and that every market answers. They do not establish where the boundary sits between question types, which would need many more pairs.

Nothing here is causal, and the mechanism is inferred. We observe that the results page carries a local pack instead of an overview. We did not run an intervention, and Google has not said why.

One surface pair. ChatGPT, Gemini and Perplexity were not measured here and may behave completely differently.

Common Questions About AI Overviews and Local Queries

Does Google never show an AI Overview for local searches?

That is not what this measured. A local question that asks for information returned an AI Overview on every run in this study. What returned nothing, in eight questions across three countries, was the local question that asks which business to choose, with or without the word “best”. The distinction is what the question wants, not whether it names a place.

Could this just be your measurement tool failing?

It was checked two ways for that reason. A control question, equally local and in the same market on the same day, returned an AI Overview through the same tool on both runs. And the live search results page for one of the commercial questions was fetched separately: it carries six local pack blocks and no AI Overview element, while the control’s page carries one in first position.

If AI Overview does not answer, where does the answer come from?

AI Mode answered all eight questions, twice each, and cited between four and fourteen domains. On this evidence that is the surface worth measuring for local buying questions, and it is a different surface with a different source set.

What should I do if my visibility report shows zero AI Overview citations?

Ask how many AI Overviews the report saw for your query. A zero out of zero is a fact about Google’s layout, not about your site, and content will not change it. A zero out of twenty is a fact about your site and worth working on. Most reports do not separate the two.

Does this mean local SEO matters more than AI visibility?

It means the local pack is still where Google answers this question, so the classic local signals still decide it, and it also means AI Mode is answering the same question with a set of sources that is not the local pack. Both are true at once, which is the uncomfortable version and the accurate one.

Will this change?

Probably. Google moves these blocks around, and a single day’s layout is not a permanent state of the world. That is an argument for measuring your own queries on a schedule rather than for taking this study as a standing fact.

Where this leaves you

The useful habit here is small and almost nobody has it: before asking whether you are cited, check whether anyone is.

It costs one query. If the surface returns nothing for your most valuable question, then every hour spent optimising for it is spent on a page that was never going to be read, and the real competition is happening somewhere you are not looking.