ktau ← Research
Research

Your brand is on the web. That doesn't mean the model ever read it.

Empirical

November 19, 2026

10

min read

ktau research team

Open access · data

A community fiber network in Vermont has its own website, coverage maps and customer reviews. Ask a leading AI model to name internet providers and it names Xfinity three times out of four — and the local provider never. We followed 248 brands from the open web, through the filtering that turns the web into AI training data, into the models and about 179,000 AI answers. The brands that get recommended are not the ones mentioned most. They are the ones mentioned in the right context — and whose pages make it into what the model learns from.

Summary

When a buyer asks an AI engine for an internet provider or a proposal-software vendor, they get a short list: three to five names out of hundreds. Most of the advice on getting onto that list is about the pages an engine finds at question time. But engines name brands both from what they find and from what the model already knows. That raises a question page-level optimisation never asks: what, in a brand's history on the web, decides whether the model knows it — and knows it as something to recommend?

We studied two very different markets. US internet providers: a few national giants, dozens of regional networks and hundreds of local ones. And proposal / RFP software: a niche B2B market where almost every vendor is small. For 248 brands we measured how they appear in the text AI models learn from, what happens to that text as training data is assembled, what a range of AI models say about them without search, and how often ChatGPT, Gemini and Perplexity name them in real answers.

Six findings:

  1. Being mentioned is not the same as being known for something. How often a vendor's name appears says little about how highly AI models rank it. How often it appears in the context of its market says much more. For the typical proposal vendor, only 2% of its mentions are in that context at all.
  2. A familiar name is not a recommended name. Seeing a name often makes a model know it. Seeing it in the right context makes the model recommend it. Models learn the first early and the second later.
  3. One word in a name, and one word in the question, move the ranking. Add "Fiber" to a cable provider's name and it climbs. Ask which provider the model would "recommend to a friend" instead of which is "best", and some models swing from national giants towards local co-operatives.
  4. Small brands can vanish from training data entirely — without being treated unfairly. The filtering is even-handed page by page. But a small brand enters it with a handful of pages, and about half the time none survives.
  5. Search does not wash the memory out. With live search on, brands the model knew in context are still two to almost five times as likely to be named, even when the engine found the same brand's website. And the brands a model already knows are the ones it searches for.
  6. New brands pay a penalty that recent visibility doesn't cover. Proposal vendors launched since 2021 rank well below established vendors with the same presence on today's web. Search narrows that gap but does not close it.

What we looked at

We built the list of brands from search results rather than from AI answers, so it includes the brands AI never mentions: 162 internet providers — national, regional and local — and 86 proposal-software vendors. For each brand we looked at how it appears in large collections of the kind of text AI models are trained on, from the raw web to finished training data, and at the words around each mention. We followed the same text through the kind of filtering used to assemble training data, asked a range of open and commercial AI models about these brands with search switched off, and compared all of it with about 179,000 answers from ChatGPT, Gemini and Perplexity, including the searches the engines ran.

Finding 1 — Count the context, not the name

The intuitive measure of a brand's presence in AI training data is how often its name occurs. For proposal software, it barely works. Many proposal-software names are ordinary words or shared with unrelated products, so for the typical vendor only 2% of its name's appearances are anywhere near words like proposal, RFP, bid or tender. Counting the name mostly measures something else.

Count only the mentions that place the brand in its market and the picture changes: every commercial model we tested ranks vendors in line with that measure, and so do the open models. Knowing a vendor's in-context mentions explains far more of its ranking than knowing how often its name appears; adding the raw count on top adds very little.

Internet providers show the same pattern less dramatically, because national providers are mentioned so often in every context. Even comparing providers of the same size — national with national, local with local — those mentioned more often in the context of internet service rank higher.

Finding 2 — A familiar name is not a recommended name

There are two different things a model can learn about a brand: that the name exists, and that it is the kind of thing to recommend. The first follows how often the name appears. The second follows how often it appears in the right context. A name that is everywhere is familiar to the model whatever you ask — so it gains little when you ask for a recommendation.

Watching a model during training shows the two developing separately. Familiar names are familiar almost from the start. The ordering of which brands the model favours when asked for a recommendation only settles later in training. The model learns who exists first and who to recommend later.

Context about a specific need matters too, when the words are right. Vendors mentioned more often alongside government contracting move up when the question asks for government-proposal software. Internet providers mentioned alongside rural and remote service are favoured for rural recommendations. But the association has to describe the need itself: loosely related words gave weak or even misleading signals.

Finding 3 — One word in the name, one word in the question

Ask the same models to rank the same providers two ways — which is "best", and which would you "recommend to a friend" — and proposal-software rankings barely move. Internet-provider rankings move a lot. Most of the models we tested push large cable and satellite providers down and small fiber providers up under the second wording; one takes Xfinity from near the top to below average.

Controlled experiments separate two causes. First, names carry cues. Adding "Fiber" to a cable or DSL provider's name moves it up the list; removing it from a fiber provider's name moves it down. Made-up providers ending in "Fiber" beat the same names ending in "Internet" or "Cable". The model rewards a technology word in a name with no evidence about the service.

Second, the question changes the criteria. Given only anonymous descriptions — technology, size, ownership — some models asked for a recommendation to a friend sharply discount national scale and switch from favouring investor-owned providers to favouring co-operatives. Technology, footprint and ownership explain most of the reshuffle. A rank in an AI answer reflects which cues the model responds to under that particular question. It is not a measure of service quality.

Finding 4 — Small brands vanish from training data, fairly

AI training data is not the web; it is what is left after aggressive filtering. Following brand mentions through that filtering, the early steps remove most of everyone's text, national and local alike. The final quality step is where the groups separate: the typical national provider keeps about 1% of its original mentions, the typical local provider 0.06% or less. Counting only mentions that place the brand in its market, the typical local provider — and the typical proposal vendor — keeps none.

The filtering is not biased against small brands' pages. It judges how a page is written: navigation and boilerplate count against it, clear explanatory writing counts for it, and that holds equally for local and national pages. Swap a local brand's name for a national one on the same pages and almost nothing changes. Page for page, local pages survive as often as national ones.

The difference is arithmetic. The typical local provider reaches the final filter with 3 to 9 pages; the typical national provider with hundreds or thousands. Keep the same fraction of each and the national brand keeps hundreds of pages while the local one, about half the time, keeps none. Fair treatment of pages does not mean fair outcomes for brands.

That is what happened to the Vermont provider in our opening example. Of 25 pages naming it in our sample of the web — mostly its own — three survive the early filtering and none survives the last step. It is absent from all 1,462 live internet-provider answers we collected, including the 1,174 in which the engine searched the web.

And what survives matters to the model. Among brands mentioned equally often on the open web, those with more mentions on pages that survive the filtering rank higher. For proposal software, the raw count of mentions on the web says essentially nothing about the models' rankings; the count on surviving pages does.

Finding 5 — Search doesn't wash the memory out

If live search simply replaced what the model already knows, a brand's presence in training data would stop mattering once the engine searches. It doesn't. In answers that used search, brands with more in-context exposure were two to almost five times as likely to be named — in both engines and both markets — even after accounting for whether the engine found the brand's own website. Being found matters too; both count.

The memory also acts before the search. The brands a model names with search switched off are the brands it writes into its search queries, in every engine and market we looked at. The brands a model already knows are the ones it goes looking for — and a brand in the query shapes which pages come back.

Across 90 days of collected answers, a model's no-search preferences predict how often ChatGPT, Gemini and Perplexity name each brand, even among brands whose websites are cited at similar rates. The gap at the bottom is stark: an individual local internet provider is named in about 0.04% of answers or less; a national one, in several percent.

Finding 6 — New brands pay a penalty that recent visibility doesn't cover

Each model's sense of a market is anchored in time: older models agree best with the web of their training years, and current commercial models with the web of recent years.

Proposal vendors launched in 2021 or later rank a quarter to a third of the market below established vendors with the same presence on today's web. Their recent visibility has not reached the models. Search helps: newcomers' websites are cited more than their web presence would predict, and their overall naming rate catches up. But when the engine has found their website, newcomers are still named at about a quarter of the odds of established vendors.

New internet providers looked like an exception — they rank above comparable incumbents. Much of that turns out to be the name: nearly half of them have "Fiber" in it, against 15% of incumbents, and several are rebranded older networks. Account for both and the advantage largely fades.

The text the models learned from also explains proposal rankings better than a vendor's current prominence does. Links, search traffic and Wikipedia readership add little once a vendor's past presence in training text is known — while that past presence still adds a lot after current prominence is taken into account. With search on, both matter about equally.

What this means for a brand

  • Volume isn't the target. Context is. A brand mentioned a thousand times in the wrong contexts is less recommendable than one mentioned a hundred times as what it is. This is sharpest in niche markets and for names that are also ordinary words.
  • The text has to make it into training data, not just exist. For a small brand, a handful of pages is a single point of failure. The filtering rewards clear, explanatory writing and penalises boilerplate — and a brand with few pages can lose all of them even when each page has a fair chance.
  • Name the need, not just the category. Being associated with the specific need a buyer asks about — government work, rural service, a particular kind of firm — moves the ranking for that need. Loosely related wording doesn't.
  • The question wording is part of the result. "Best" and "recommend to a friend" can produce different winners. Measuring a brand's standing means measuring it under the questions buyers actually ask.
  • Being found won't fix not being known. The model's memory decides which brands enter the search and still predicts who is named after it. Being found and being known have to be built together, and checked on the live engines over time.

Limitations

These are associations, not controlled interventions on training. The training data of commercial models is not public, and their answers also reflect later tuning, so their link to the text we measured cannot be attributed to training text alone. A brand's own website is only part of what an engine finds when it searches; what third-party pages say was not measured. Our measures of context are deliberately simple and can miss relevant text or count irrelevant text. The study covers two US markets, and the models change continuously; the measurements describe 2026.


Want to know whether the models know your brand — and as what? Talk to the ktau team.