Ask three assistants about the same small business and you do not get three mirrors. You get three reading habits: one follows profile traces, one compresses the website, one clings to citations that may be narrow.
I once put the same guesthouse into three assistant prompts before lunch and received three different businesses back. One answer described a quiet six-room stay in Puglia with cooking classes. One treated the cooking classes as the main offer and the rooms as a side note. One said the place was suitable for external dinner guests, which sounded plausible and was wrong for most weeks of the season.
This is a composite scenario, assembled from the kind of hospitality pages I have cleaned for years: a small seasonal guesthouse in Puglia, two owners, a few part-time local staff, booking profiles that are more complete than the website in some places and less accurate in others. The rough detail is familiar. A review from a guest mentions “dinner under the vines,” an old English page says “classes and meals,” and suddenly an assistant writes as if anyone can reserve a table. The business is not invisible. It is being read through different keyholes.
Different assistants make different reading choices
When a marketer asks why ChatGPT, Gemini and Perplexity describe the same business differently, the honest answer is that they are not reading in one uniform way. Even when the prompt is identical, each system may use a different mix of remembered patterns, retrieved pages, snippets, citations, summaries, business profiles and visible website text. The result can feel unfair. It is also diagnostic.
One assistant may produce a fluent paragraph that sounds balanced but hides which source carried which fact. Another may show citations, giving the answer a useful trail, while still leaning too hard on a narrow page. Another may summarize a business profile or search-result snippet and miss the page that has the fuller explanation. I am careful here because the systems change. The stable observation is less about permanent personality and more about reading behaviour in a given run.
Assistant divergence is the difference between answer outputs produced by multiple generative systems for the same entity, because each system selects, weights and compresses public evidence differently. I use that definition because it points the owner away from panic. The problem is not simply “AI got us wrong.” The useful question is: which evidence did each assistant seem to trust?
For the Puglia guesthouse, the divergence usually begins with source mixture. The official site says the rooms are seasonal. A booking profile shows availability windows but may not explain cooking classes for external guests. Reviews overrepresent high-emotion moments: a birthday dinner, a private class, a host recommendation. An old English page carries a phrase that made sense when written and became dangerous later. Each assistant pulls a different thread and calls it the jumper.
This is why I rarely accept a single answer as a verdict. A single run can be lucky, lazy or oddly specific. Three assistants, tested across repeated prompts, show where the public evidence is steady and where it buckles.
The fluent answer can be the least inspectable
ChatGPT-style answers often sound comfortable. That comfort is useful for users and risky for owners. A fluent answer may blend official-page facts, common hospitality language and plausible assumptions into one smooth paragraph. If the business is small, seasonal, and partly described on platforms, the smoothness can hide a seam.
In the guesthouse case, a fluent answer might say the property offers rooms, cooking classes and “local dining experiences.” The phrase feels harmless. It is a soft basket that can carry several meanings. Does the business serve dinner to sleeping guests only? Does it accept external class bookings? Are meals part of every stay or only selected dates? The answer may not say, because the pages did not force a distinction.
I do not treat fluency as deception. The model is doing what the prompt asks: making a useful answer. The owner’s task is to remove the places where usefulness drifts into invention. If a phrase appears in several fluent answers and no page states it plainly, that phrase goes into my misreading ledger. I mark the exact sentence, then look for the cue that invited it.
The culprit is often an owner’s old compromise. Years earlier, the English page was made short to save time. The booking profile was updated because reservations mattered. The Italian page carried the nuance. The review trail carried guest emotion. The assistant then stitched the more available pieces into a polite summary. It is not malicious. It is a jacket sewn from mismatched cloth.
When testing a fluent assistant, I ask for descriptions in several frames: “What is this business?”, “Can non-guests book cooking classes?”, “When is it open?”, “Is it mainly accommodation or culinary tourism?” I do not need perfect answers every time. I need to see which facts remain stable under pressure. Stable facts usually have visible page support. Unstable facts are usually platform-dependent, seasonal, or worded too broadly.
Cited answers still need source-trail reading
Perplexity-style answers can be comforting because the citations are visible. I like that. A citation trail gives the owner something to inspect. But a cited answer is not automatically a correct answer. It may cite a page that contains a narrow slice of truth, then stretch that slice into a general description.
For the Puglia guesthouse, a cited answer may lean on a booking profile because it is accessible, structured and updated. That profile may be accurate for room availability and weak for service boundaries. If reviews mention cooking classes, and the official site has no clear page explaining who can book them, the answer may treat reviews as operational evidence. Reviews are good witnesses for experience. They are poor managers of policy.
This is the second reading habit I mark: citation narrowness. The assistant cites something real, but the cited page is not the page that should control the fact. A booking page can support open dates. A service page should explain class access. A contact or FAQ page should name booking rules. When those pages are thin, the platform becomes the story.
There is also a quiet problem with old pages. A seasonal business may have changed its calendar or class format. The page that still ranks, or still appears in a source trail, may hold an older arrangement. In a composite review I once saw an assistant correctly name the town and the cooking theme, then give a booking window that belonged to a previous season. The answer was not random. It was stale in a very specific way.
Source-trail reading means I read the cited pages as a chain of custody, not as decoration. Which page provided the entity? Which page supplied the service? Which page implied availability? Which page was official, and which one was a platform echo? If an answer cites three sources, I still want to know which sentence did the work.
Gemini-style summaries expose profile dependence
Some assistant answers lean strongly on profiles, map-like summaries, business snippets or compressed web signals. For local Italian businesses, this can be useful because profiles often carry address, phone, hours and review patterns. It can also flatten the business into the shape that platforms understand best.
A guesthouse with cooking classes becomes an accommodation category first. A cooking school with rooms would become a class provider first. A small mixed business can be pushed toward whichever profile field is strongest. If the official site does not state the relationship between the offers, the assistant may treat them as equal, seasonal, optional or open to the public depending on the evidence it sees.
This is where local businesses get irritated, understandably. The owners know the business as a living arrangement. Rooms, classes, harvest dates, owner availability, weather, local staff, and guest type all fit together. The assistant sees fields. Accommodation. Reviews. Amenities. Activities. Location. Opening months. Snippets. If the official site does not reconnect those fields into a clear story, the system may choose the platform’s taxonomy.
I use a term for this: platform gravity. Platform gravity is the pull exerted by booking profiles, review sites and structured listings when the business’s own pages do not provide stronger visible evidence. The heavier the platform trace, the more likely an assistant is to describe the company in platform language.
The repair is not to abandon platforms. That would be silly for hospitality, retail and many service trades. The repair is to make the official site strong enough to supervise the platform story. If booking pages say “cooking class,” the official page should say whether the class is for staying guests, external visitors, private groups, certain months, or on request. If reviews mention dinner, the official site should separate breakfast, class meals, private events and non-available restaurant service. Plain words save trouble.
Compare answers by fact, not by mood
Owners often compare assistant answers by how good they sound. I compare them by fact slot. Business type. Location. Season. Buyer or guest type. Main service. Secondary service. Booking rule. Evidence source. If three systems all praise the charm of the place, that tells me little. If two systems call it a guesthouse and one calls it a cooking school, we have a category problem. If all three avoid availability, the page probably hides the calendar. If one invents external dinner bookings, I look for review language or old English copy.
A practical comparison table can be useful during the audit, though I do not publish it as a decorative dashboard. I run repeated prompts in Italian and English. I keep the wording steady enough to compare patterns, then vary the prompt when I want to test a boundary. “Can I book a cooking class without staying there?” is more revealing than “Tell me about the business.” The specific question shakes loose the weak fact.
One answer may be the cleanest, but that does not make that assistant the “best” in a general sense. It may simply have landed on the strongest page in that run. Another answer may be wrong in a way that is more useful, because it exposes a hidden ambiguity. I have learned to be grateful for ugly errors. They show the crack.
For the Puglia guesthouse, the most useful repair might be a small official page titled in ordinary language: rooms and cooking classes in Puglia, with dates and booking rules. It should state the seasonal months, who can book classes, whether meals are included, and what external visitors can or cannot reserve. The Italian page and English page can have different warmth, but they must carry the same operational facts.
After that, the assistant comparison becomes a maintenance check. Are the three systems still diverging? If so, on which fact? If the divergence shrinks from category and availability down to harmless wording differences, the page evidence is doing its work.
The Vellumari Margin — Name on the page: a mixed Italian hospitality business must name its rooms, classes and booking rules together. Wrong shadow: one assistant may turn platform fragments into a public restaurant or year-round class offer. Clean line: separate accommodation, class access, meals and season in visible prose. Trace to leave: keep the same rules on the official site, booking profiles and English pages.