Signal · AI visibility
Four AI engines can name four different local winners
Appearing in one AI assistant is not proof that you appear in the others. And looking once is not measuring.
September 24, 2026
The short version
Two provider studies found that different AI assistants frequently name different local businesses for the same question, and that asking the same question again often returns a different list. A screenshot from one assistant on one day tells you about that assistant on that day, and nothing more.
Illustration, not a case study
The check that felt like proof
A family group runs six trattorias in one city. On a Tuesday an area manager opens an AI assistant on their phone and types "best pasta near me". The Kruiskade location comes back first. They screenshot it, post it in the team chat with "we are winning AI", and everyone moves on.
Two weeks later a franchise partner asks the same question in a different AI assistant, from the same street, and the group is not mentioned at all. Nothing broke in between. No listing was deleted, no review was lost, no website went down.
Both observations were true. Each one described one assistant, at one moment, from one place. Neither described the business.
An AI assistant — also called an AI engine — is a product that reads a question written in ordinary words and answers in ordinary words, often naming a small number of businesses directly. Some also show links, and some show nothing else. What matters here is that the short list of named businesses is assembled by the assistant, and that there is now more than one assistant, and they do not share an answer.
The trattoria group, the two checks and the team chat are invented, written to show the shape of the problem. They are not a customer, a case study or a measurement. The measured figures on this page are in the two boards below, each beside its source.
What changed
Several assistants now answer the same local question independently, each assembling its own short list. Two providers have published counts of how far apart those lists sit, and the gap is not a rounding error.
The consequence for an operator is narrow and practical: one spot-check can no longer stand in for the whole picture, and no single number can be bought, won or reported as "our AI rank".
Measured — agreement between engines
How often four engines put the same business first
- local cases where all four engines answered the same questionSample size
- 491,972
- of those cases where all four named the same business firstAgreement
- 4%
- of those cases where at least one of the four put a different business firstThe arithmetic complement of the 4% above, not a separate measurement.The remainder
- 96%
Four-way agreement on the top business was the exception in this sample, not the rule.
SourceYext — "Four AIs, one question, four different answers"Study window February–August 2026. The primary page is dated September 2026 and gives no day. Source checked 22 September 2026.
Measured — stability of one engine over repeat runs
What happens when you ask the same question again
- localized searches analysedSample size
- 200,085
- average overlap between the businesses named on one run and on a repeat runOverlap: write down the businesses named each time you ask; overlap is the share that appears on both lists.Overlap
- 20–33%
- of businesses named in one run were still named in a later runPersistence: whether a business that was named once is named again later.Persistence
- ~50%
The instability is not only between engines. It is inside a single engine, between one run and the next.
SourceBrightLocal — "What 200k local AI searches tell us about what businesses get recommended, and why your website matters"Published 16 September 2026. Source checked 22 September 2026.
Illustration — the operator’s mental model
One question, four independent answers
One question, asked from one street
"Best pasta near me"
Four assistants answer separately
| Assistant | Named first | Where the trattoria appears |
|---|---|---|
| Assistant 1 | Trattoria Vesta — Kruiskade | First of four named |
| Assistant 2 | A competitor two streets away | Third of five named |
| Assistant 3 | A competitor in another district | Not named |
| Assistant 4 | Trattoria Vesta — Kruiskade | First of three named |
| Same business first in all four | No. Two assistants open with the trattoria, one opens with a competitor and mentions it late, one does not mention it. The manager screenshotted Assistant 1. | |
One look is a sample, not a measurement
A spot-check answers a small question: did this assistant name us, once, from this place, at this moment? That is a single draw from something that moves.
A measurement answers a larger one: how often are we named, across which assistants, asked from which places, over which weeks? The unit is not the screenshot. It is the dated, repeated observation — the same questions, the same places, on a schedule, written down.
Repeated weekly runs will tell you more than forty screenshots taken on one afternoon. How many weeks is a judgement, not a finding: the cadence suggested further down is Obenan’s starting point, not a threshold either study established.
What the evidence supports
Being named, and what it is not
Being named tells you
- This assistant was willing to mention you for this question, from this place, at this moment.
- You entered the consideration set — the short list the assistant is prepared to put in front of someone.
Being named does not tell you
- That you were chosen. Entering a short list and being selected from it are two different events.
- That the other assistants named you, or will name you next week.
- That anyone walked in, booked a table, paid, or was served.
- That there is a rank to win. There is no single leaderboard shared across assistants.
- That your name, address, hours and category were read correctly. A mention is not a check on your own facts.
Keep these two columns in two different reports. Most of the confusion in this area comes from one being read as the other.
Related reading
The distinction in the right-hand column above — appearing on a list, and being the one picked from it — is the subject of an earlier Briefing.
Keep the stages apart
Seven separate events, seven separate records
An AI mention sits at the very first of these. Reporting it as any of the later ones will overstate what happened, and will make the later drop look like a failure that never occurred.
- Organic discoverySomeone is shown your business without anyone paying for that placement — including an unpaid mention by an AI assistant.
- Paid exposureA placement you bought. It is a different budget line and a different record, even when it appears on the same screen.
- ReferralAnother site, partner or platform sends the person onward to you.
- LeadThe person identifies themselves — a call, a form, a booking request, an enquiry about a table.
- Customer approvalThe person agrees to the specific thing: this table, this time, this order.
- PaymentMoney is actually captured. Approval and payment are not the same event, and either can fail without the other.
- FulfilmentThe meal is served, or the order is handed over.
Which of these you already record varies by business. In most, the later stages leave a trace somewhere — a booking system, a till, a kitchen ticket — while the first one is the one nobody is writing down.
Practical next steps
A starter cadence for the next four weeks
None of this requires a new system. It requires writing things down on a schedule, and keeping the stages apart. The four weeks are Obenan’s suggested starter cadence — a practical minimum for seeing a pattern rather than a run — and not a threshold either study established.
- Write down the five questions a real customer would askPer location, in the language that location is actually searched in. Plain wording, not marketing wording — "best pasta near me", not "premium Italian dining Rotterdam".
- Ask each question in each assistant you care aboutFrom the location’s own city, not from head office. Record the date, the time, the assistant, the question and the place you asked from. Those five fields are a basic log — enough to start writing runs down, and not a measurement system on its own.
- Record two different things on every runWere you named at all, and were you named first. These move independently, and collapsing them into one number hides most of what is happening.
- Repeat weekly for at least four weeks before concluding anythingOne run tells you about that run. Four weekly runs give you a first dated baseline — a record your later runs can be compared against. It is a starting point, not a verdict on how visible you are. Four weeks is the cadence Obenan suggests starting with; adjust it once your own record shows how much your results move.
- Keep your own published facts consistent, per locationName, address, phone, opening hours, category, menu or services — identical everywhere you publish them. This covers the sources you control; neither study establishes what any assistant reads. Contradictions between your own published sources are one thing you can act on directly, whatever any assistant then does with them.
- Keep discovery, leads, payments and fulfilment in separate columnsNever let an AI mention be reported as revenue. When the mention improves and bookings do not, you want to be able to see that clearly rather than argue about it.
- Compare against your own baseline, not a provider averageThe published ranges describe the samples those providers ran. Your own record describes you. After any change, compare your new weeks to your old weeks.
Evidence boundary
What this evidence does not prove
These are provider studies with their own question sets, their own engines, their own places and their own dates. Read them as evidence that the disagreement is real and large, not as a forecast for your market.
- They do not prove a result for every market, every category or every business, including yours.
- The 4% is how often four engines agreed on one top business in that sample, over that study window. It is not a probability that applies to a given business.
- The overlap and persistence ranges are averages across a large sample. Your own numbers may sit above or below them.
- Nothing here shows that one assistant is better, more accurate or more worth optimising for than another.
- Nothing here shows a universal rank exists, or that any action produces one.
- Nothing here measures customers, bookings, payments or revenue. Entering a consideration set is not being selected, and being selected is not being paid.
- The four-week cadence suggested above is Obenan’s practical starting point. Neither study tested it or established it as a threshold.
Sources
Every figure above is attributed to one of these two entries, beside the figure itself.
Yext — "Four AIs, one question, four different answers"
Source of the 491,972 shared local cases and the 4% four-engine agreement on the top business.
Study window February–August 2026. The primary page is dated September 2026 and gives no day. Source checked 22 September 2026.
Source of the 200,085 localized searches, the 20–33% average overlap on repeat runs and the roughly 50% persistence.
Published 16 September 2026. Source checked 22 September 2026.