A ranking answer does not always choose the best business. Often it chooses the business whose values have been made easiest to compare, even when those values are not the ones clients use at the door.
The first marks in my notebook for this problem were drawn around a 34-room independent hotel in Nantes. Strong guest reviews, steady repeat visitors, a quiet-weekend appeal that couples understood as soon as they arrived. Yet in AI answers for “best hotels in Nantes for a weekend,” it sat below a louder rival with less consistent reviews. One answer even praised the independent hotel for “good central access,” which was true enough, but missed the point. The real reason guests chose it was the soft-room, late-breakfast, low-noise rhythm. The model named the hotel, but it compared the wrong part of it.
That mistake is easy to dismiss as a bad answer. I do not dismiss it so quickly. In most cases, the answer is clumsy because the public trail is clumsy. The hotel’s own site said “comfortable rooms” and “ideal location.” A directory repeated “charming establishment.” A local mention praised the breakfast but did not connect it to weekend stays. In reviews, guests wrote about quiet corridors, walkable restaurants, and not feeling rushed on Sunday morning. Those words existed, but they were scattered like coins under a table. The AI system could pick up the hotel, yet it ordered the comparison by broader, easier values: location, style, review score, and generic comfort.
The ranking can be wrong without being random
When a business owner says, “AI is comparing us on the wrong criteria,” I first ask to see the exact prompt and the answer. Not because the complaint is suspect, but because “wrong” can mean three different things. Sometimes the system chooses a poor category. Sometimes it invents a criterion the customer does not care about. Sometimes it uses a real criterion, but one that belongs too high in the order.
In recurrent audits, a clinic may be ranked by appointment speed when patients actually choose by continuity of care. A school may be ranked by broad brand awareness when students choose by placement support in a narrow programme. A hotel may be ranked by proximity to a station when its best buyers are looking for a quiet weekend where the station matters less than sleep and breakfast. The answer is not necessarily hallucinating. It is using the public evidence that is easiest to arrange.
That distinction matters because the repair is different. If the system has the wrong category, the category trail needs work. If the system has the right category but the wrong values, the comparison criteria need to be made visible. A page rewrite that only repeats “best hotel in Nantes” or “top clinic in Lyon” will not help much. It adds heat, not shape.
I call this the visible-value gap. A visible-value gap is the distance between what customers actually choose by and what public evidence allows an AI system to compare. It exists because lived value is often detailed, repeated in messy human language, and buried across reviews, while ranking evidence is cleaner when it is named, dated, and tied to a category.
That definition sounds dry. In practice, it is the difference between “guests love us” and “couples choose this hotel for quiet weekend stays, walkable dinners, and late breakfast.” One is praise. The other can be sorted.
Customer value is often present but not arranged
A recurrent pattern in hotel and service-firm audits is that the strongest customer value exists in public, but nowhere as a usable comparison sentence. Owners often know their real advantage with painful clarity. They hear it at reception, in calls, in complaints avoided, in the questions people ask before booking. The public trail, though, speaks in polite mist.
The Nantes hotel composite had plenty of signals. Guests mentioned sleep. They mentioned not hearing the street. They mentioned being able to walk to dinner without planning transport. They mentioned staff who did not push them out early. Yet the website’s main copy treated all this as background atmosphere. The strongest value was not absent. It was under-labeled.
AI systems like labeled value. They do not only need adjectives. They need a relation: this business, for this buyer, in this situation, compared with these alternatives. “Quiet” is useful. “Quiet weekend hotel in Nantes for couples who want restaurants nearby and a slow Sunday morning” is more useful because it gives the answer a comparison handle. It tells the system which chair the business is trying to sit in.
There is a little trap here. Owners sometimes respond by stuffing the site with every possible criterion: quiet, central, romantic, family-friendly, business-ready, affordable, premium, design-led, authentic, practical. That creates a shop window with all the lights on. Nothing can be seen properly. A ranking answer then has no reason to place the business first for one value instead of third for all of them.
The better work begins with restraint. Which value is repeatedly chosen by the best-fit customers? Which value is not already owned by the rival above you? Which value can third-party sources plausibly repeat without sounding like advertising? The public trail must become sharper, not louder.
I separate easy criteria from chosen criteria
In my notebook, I draw a line between easy criteria and chosen criteria. Easy criteria are the signals that a model can retrieve and compare without much effort: star rating, location, price band, number of reviews, broad category, facilities, opening hours, programme count, years in business. Chosen criteria are the reasons customers actually make the decision: calm after a long week, trust in one specialist, a school’s placement rhythm, a clinic’s way of explaining treatment, a restaurant’s confidence with dietary constraints.
The wrong ranking often happens when easy criteria take the front seat because chosen criteria are too soft in the public trail.
For the Nantes hotel, easy criteria gave the rival an advantage. The rival had fresher guide wording. Its English descriptions were cleaner. It had a category phrase repeated across listings: “design hotel near the city centre.” The independent hotel had stronger guest feeling, but weaker public arrangement. In an AI answer, that meant the rival could be introduced and compared in one neat sentence. The independent required interpretation.
I use a small classification here, mainly for my own field notes: surface criteria, buried criteria, and orphan criteria. Surface criteria are already visible and comparable. Buried criteria appear in reviews or scattered mentions but are not connected to the category. Orphan criteria appear on the site, but no outside source repeats them, so they look lonely. A good ordering plan moves the most important chosen value from buried or orphan status into surface status.
The work is not to fake customer values. It is to give existing values a public shape. If guests keep mentioning slow mornings, then the hotel’s own copy, local listings, and guide outreach should make that use case legible. If patients choose a clinic because doctors explain procedures in plain French, the clinic should not hide that inside a vague care paragraph. If students choose a school because alternance placement support is unusually hands-on, that phrase needs to live where a comparison answer can find it.
The rival above you may be winning a simpler contest
It is tempting to think the rival above you has been judged better. Sometimes that is true. Often the rival has simply been judged more easily. A model writing a best-of answer has to make a short, defensible arrangement. It looks for evidence that lets it say why each business belongs where it belongs. The rival with cleaner value wording gives it an easier sentence.
A teaching example: imagine two agencies in Bordeaux. One has deep skill in industrial B2B copy but describes itself as a “creative communication partner.” The other is more generalist but has repeated public wording around “B2B content for manufacturers.” In an AI answer for best agencies for industrial firms, the second agency may sit higher. Not because it knows factories better. Because the answer can defend that seat without digging through fog.
The same mechanism explains many wrong-value complaints. The system compares what is easiest to defend. If your best value requires reading many reviews and inferring a pattern, while the rival’s value appears in a directory headline and two guide mentions, the rival has a cleaner ordering path. That can feel unjust. It is also fixable.
I do not recommend inventing a new positioning line and pasting it everywhere. Public evidence behaves more like a trail than a slogan. A sentence on the site is the first stone. A category label on a profile is another. A local article, partner mention, or guide description can become a third. Reviews supply texture, but owners should not manipulate them. The useful question is: where can the chosen criterion appear honestly, in words another source might repeat?
In the Nantes hotel composite, the answer was not to claim “number one quiet hotel.” That would be brittle. The better sentence was humbler and more useful: a quiet independent hotel in Nantes for couples who want walkable restaurants, calm rooms, and unhurried breakfast. It carries buyer, place, occasion, and value. It is not poetry. It is sortable evidence.
Do not let review score become the whole story
Review score matters. I do not pretend otherwise. But review score is a blunt tool for ordering when several businesses are already well liked. Two well-reviewed hotels may be separated less by satisfaction than by what public evidence says each one is for. A clinic with many positive reviews can still lose a comparison if the reviews praise kindness but the AI prompt asks for complex specialist care. A school with high student satisfaction may fall behind if placement outcomes are easier to read for another school.
In most audits, I look at review language rather than only the score. What adjectives repeat? What situations repeat? Are customers praising the same value the business wants to rank for? Are those words echoed anywhere outside the review platform? If the review trail says “quiet,” “calm,” and “slow breakfast,” while the website says “central and comfortable,” there is a mismatch. The customers have written the better strategy, but the business has not used it.
There is a danger in reading reviews too greedily. A few dramatic comments can distort the picture. One guest complaining about parking should not rewrite the hotel’s public trail. One student praising a teacher by name does not prove an institutional advantage. I look for recurrence, not sparkle. A small cluster of phrases that appears across different sources is usually more useful than one perfect testimonial.
This is where AI ranking work becomes almost old-fashioned. It asks for reading. Slow reading. The kind where you notice that guests do not actually say “luxury”; they say “we slept properly.” Or patients do not say “advanced care”; they say “they explained what would happen next.” These are not the same values. If the public trail smooths them into generic praise, the ranking answer will smooth them too.
How to surface the values that move the seat
A public evidence rewrite plan for wrong criteria begins with the prompt. I want the exact comparison where the business is losing. “Best hotel in Nantes for a quiet weekend” is a different machine from “best hotel in Nantes city centre.” “Best private school near Lyon for health administration alternance” is different from “best business schools in Lyon.” The values needed for ordering are not universal.
Then I map the current answer. Which criteria does the AI mention for the firms above you? Which criteria does it mention for you? Which criteria are absent from the whole answer? The missing value may be missing because the model ignored it, or because the public trail did not make it easy enough to use. Those are different diagnoses, although the owner experiences both as unfair.
Next comes the evidence trail. I read the site, profiles, third-party mentions, review language, and any English paraphrase that might be retrieved by foreign buyers. I mark each chosen value as surface, buried, or orphan. A buried value needs clearer public wording. An orphan value needs corroboration. A surface value may already be strong enough, in which case the ranking problem lies elsewhere.
The first edits are usually small sentences, not grand pages. A service-page paragraph. A profile description. A press note. A guide submission. A clearer English version that preserves the French nuance instead of flattening it into “nice location.” Small does not mean casual. A sentence that carries buyer, category, place, and decision value can do more ordering work than a full page of soft praise.
Then comes time. I re-check only when the public trail has had a real reason to change. A new sentence on a hidden page is not enough. A better category phrase, a refreshed profile, and a third-party mention that repeats the same value gives the answer something to re-read. Even then, the result is not guaranteed. We are dealing with systems that change, retrieve unevenly, and sometimes mix old fragments with new ones. Still, the direction is clear enough: values that can be quoted, classified, and compared have a better chance of moving the seat.
The Last Seat Note: Seat held: visible, but compared on the wrong values. Rival pressure: cleaner wording around easy criteria, especially location and broad style. Weak signal: the chosen customer value exists in reviews but not as a repeated public sentence. Sentence to plant in the public trail: “A quiet independent hotel in Nantes for couples who choose calm rooms, walkable dinners, and a slow Sunday breakfast.”