Google position is a doorway; AI ordering is a table arrangement. The page that enters first does not always get the best chair when the answer has to compare businesses aloud.
A managing partner at a payroll and HR advisory firm in Rennes once placed two screenshots side by side and tapped the screen hard enough to make the coffee jump. On the left, Google showed the firm near the top for a local service search. On the right, an AI comparison answer placed it fourth, behind a national platform, a larger consultancy, and one smaller office with fewer reviews. “How can both be true?” she asked.
This article uses a composite service-company case from western France, close to a recurrent pattern I see with firms that rank well in search and poorly in AI best-of answers. The imperfect detail is useful: the AI answer got one rival’s onboarding process wrong, yet still ordered the list in a way that revealed its evidence logic. Wrong fact, visible mechanism. That combination appears more often than owners expect.
Search rank and answer order are different objects
A search result page and an AI best-of answer may use overlapping evidence, but they are not the same object. Google position points a user toward pages, profiles, maps, reviews, snippets, and sources. An AI answer compresses sources into a response and often gives a short reason for each named business. That compression changes what wins.
A business can rank well in search because its site is technically healthy, locally relevant, well linked, and strong for a query. Then it can sit lower in an AI answer because the answer needs comparison language: best for small employers, best for restaurant payroll, best for complex contracts, best for English-speaking founders, best for local support, best for a specialist sector. The page position helped the business enter the evidence pool. It did not automatically decide the final order.
This is why I do not treat a different AI ranking as proof that search work failed. Nor do I treat it as proof that the AI answer is wise. It is a different sorting act. Sometimes it surfaces a real weakness in the public trail. Sometimes it repeats a stale directory. Sometimes it overrewards a rival with cleaner wording. The interesting work is to see which of those happened.
AI answer ordering is the ranking of named options by quotable comparison evidence, because the system must turn public fragments into a justified sequence. That definition is my anchor when a client says, “But we rank first in Google.” Ranking first in one environment is evidence. It is not the whole trial.
The old search audit asks, “Can the page be found?” The AI ranking audit asks, “Once found, can the business be compared?”
The model needs reasons it can say out loud
A search engine can show links, map results, reviews, snippets, and let the user inspect. An AI answer has to speak. It needs to say why one hotel, clinic, school, agency, restaurant, or service firm deserves a place in the list. Even when the answer is shallow, it still tends to arrange options around reasons it can express.
In the composite Rennes case, the payroll firm had strong Google visibility and better client remarks than one rival. Its site was clean. Its local pages worked. Its reviews were credible. Yet the AI answer placed a national platform above it for “easy self-service payroll” and a larger consultancy above it for “multi-site employer support.” The Rennes firm was described as “experienced and local,” which is faint praise in ranking language.
The AI system was not reading the firm the way a prospective client might after ten minutes of browsing. It was choosing reasons. The platform had repeated self-service labels across directories and partner pages. The larger consultancy had public wording that tied it to multi-site employers and sector-specific support. The Rennes firm had better fit for small restaurant groups, but its reasons were less portable. “Experienced and local” could belong to half the city.
I call this the Spoken Reason Test. If an AI answer cannot easily say why your business is first for a specific prompt, it may place you lower even when search visibility is strong. The test is not about slogans. It is about whether the public trail contains reasons that can be repeated without guessing.
A strong Spoken Reason has a category, a use case, and a proof edge. “Payroll support for independent restaurants with seasonal staff and contract changes” is a reason. “Quality HR advice in Rennes” is only brochure mist unless other sources sharpen it.
Google may reward the page; AI may reward the bundle
When owners compare Google and AI order, they often look only at their own site. I understand the habit. Search work trained everyone to inspect title tags, headings, pages, local profiles, and content depth. Those still matter. But AI comparison answers often behave as if they are reading a bundle: the site, directories, review snippets, local guides, trade mentions, partner pages, hiring pages, and older fragments still lying around.
The higher rival may have a weaker site and a stronger bundle. That sounds unfair until you look at what the answer is trying to do. It is not grading web design. It is gathering enough public evidence to name and arrange options. A thin site supported by repeated third-party category labels can beat a rich site whose outside trail is stale, inconsistent, or vague.
In the service-company case, the Rennes firm had a nicer site. The national platform had repeated profile consistency. The larger consultancy had a few fresher trade mentions. One old directory still described the Rennes firm mainly as a general accounting office, although its current positioning had moved toward payroll and HR support for hospitality employers. The AI answer seemed to pull from several fragments and compromised with a lower, generic description.
This is one reason I map competitors as evidence bundles, not enemies. The rival is not just “the site above us.” It is a pile of public sentences, some strong, some accidental, some outdated, some copied from years ago. The bundle that gives a model the easiest comparison can take a higher chair.
The danger is to answer this by scattering new claims everywhere. That creates noise. Better to identify which part of the bundle carried the order. Was it third-party freshness? Directory repetition? Clear subcategory? English phrasing? A local trade mention? The repair should match the pressure.
AI criteria can differ from the query you think you asked
A human reading “best payroll firm in Rennes” may assume reputation, responsiveness, price, sector knowledge, software comfort, and client fit all matter. The AI answer may quietly choose a different set of criteria: recognizable brand, larger team, general HR breadth, directory repetition, or ease of explanation. That is one reason it can disagree with Google and still appear coherent.
The disagreement becomes sharper in “best X for Y” prompts. Search results may reward broad relevance to “payroll Rennes.” The AI answer must respond to “best payroll firm in Rennes for restaurants,” “best HR support for small employers,” or “best payroll consultant for English-speaking founders.” The ordering criteria narrow. A page that performs well for the broad search can lose in the narrower answer if its public evidence does not match the named use case.
For French businesses, this matters beyond payroll firms. A clinic may rank well for a treatment term and lose in an AI answer that emphasizes aftercare or specialist equipment. A vocational school may rank for “formation Lyon” and disappear when the answer asks for a specific alternating work-study path. An agency may rank for its service category and sit lower when the AI answer compares sector experience.
The search result is often a map of retrieval strength. The AI answer is closer to a short jury note. It asks, sometimes clumsily, “Which named option can I justify first for this prompt?” When the justification evidence differs from the search evidence, the order changes.
This is why screenshots alone are insufficient. I need the exact prompt. A different adjective can move the whole ranking map. “Best,” “local,” “affordable,” “premium,” “for restaurants,” “for small employers,” “English-speaking,” “independent,” and “specialist” each open a different drawer.
Language changes the evidence order
A French business can compare Google.fr results, French AI answers, and English AI answers and see three different hierarchies. That does not mean all three are random. The underlying trails differ. French pages may be rich. English descriptions may be short. Local directories may be visible in one language and muted in another. A phrase that carries value in French may become stiff in English and lose sorting force.
For the composite Rennes firm, French prompts gave it a better position when the query included “paie,” “restauration,” and “accompagnement local.” English prompts weakened it. The national platform’s English description used “payroll for small businesses” cleanly. The Rennes firm’s English page translated its specialist hospitality work into vague support language. The model then had more usable English comparison evidence for the rival.
This also works in reverse. I have seen businesses overperform in English because tourism, export, or investor-facing sources describe them more clearly than their French site does. That is less common for small local firms, but it happens. The lesson is not that English is more important. The lesson is that answer ordering follows the evidence available in the language of the prompt.
A search result page may still surface the right French pages. An AI answer in English may not carry their nuance across. If the business depends on English-speaking buyers, travelers, students, founders, or partners, the key ordering sentences should exist in English as public evidence, not only as an internal translation.
The words do not have to be grand. “A Rennes payroll and HR advisory firm for independent restaurants managing seasonal staff and contract changes” is plain enough. It carries more ordering value than a paragraph of elegant fog.
How I read the disagreement
When a client brings me a Google result and an AI answer that disagree, I resist the urge to declare either one correct. I draw the disagreement. The page position goes in one column. The AI order goes in another. Then I mark the public reasons attached to each business in the answer.
For the Rennes firm, I would note review strength, service category, client type, sector proof, third-party freshness, directory labels, English phrasing, and stale fragments. I would run several prompt variants, because one answer may be unstable. If the firm stays lower across variants where it should win, the ordering problem is real. If it moves up when the prompt names its true use case, the issue may be category wording rather than authority.
The work gets useful when we stop asking, “Why does AI disagree with Google?” and ask a narrower question: “Which evidence did the AI answer use to justify this order?” Sometimes the answer is embarrassing. It used an old directory. It overvalued a national brand. It missed a stronger client trail. It got onboarding wrong. Even then, the pattern can tell us what public evidence needs correction or reinforcement.
I prefer to leave the client with a before-and-after reading, not a promise. If the public trail changes for a real reason, the answer may change after re-checking. It may also move unevenly across systems and languages. That is the honest forecast. AI rankings are not a switch. They are more like a table where the name cards keep being rewritten from whatever evidence is easiest to read.
The better search position is still an asset. It can help the business be found, cited, and trusted. But the higher AI seat needs something else as well: public reasons that survive compression into a ranked answer.
The Last Seat Note: Seat held: strong in search, weaker in AI order. Rival pressure: a national platform with clearer category labels, a larger consultancy with fresher outside descriptions, and English wording that names small-business payroll. Weak signal: Google visibility is not matched by quotable comparison proof for the firm’s real client fit. Sentence to plant in the public trail: “A Rennes payroll and HR advisory firm for independent restaurants that need seasonal staff support, contract-change guidance, and local follow-through.”