Why your brand isn't in ChatGPT answers: known vs found

Known vs found: the two ways a brand gets into an AI answer

A brand gets into an AI answer in one of two ways: the model already knows it from training, or the engine finds it in a live search. We call these known and found. Researchers call them parametric knowledge and retrieval. The difference matters because the fixes are different and they work on different clocks.

Known means the brand is stored in the model's weights. The model learned it from the text it was trained on, and it can name the brand with search turned off. That memory is frozen until the next model release. Found means the engine ran a web search, pulled pages into its context and used them to write the answer. Found can change from week to week, because it depends on what the search returns today.

A brand can be known but not found, found but not known, both or neither. Each combination points to a different plan. The glossary entry for known vs found has the short definitions.

How often ChatGPT searches vs answers from memory in 2026

ChatGPT searches the web in a minority of all conversations but in most buying conversations. Across general use, vendor datasets put search at about 14–18% of conversations or responses (Profound, 700k conversations; Ahrefs logged-out data, reported second-hand). For commercial prompts, one small study found search in over 80% of them (cloro, 2026). These are vendor studies with large but uneven samples, so treat the figures as plausible rather than settled.

So for a broad question such as “what is a CRM”, ChatGPT mostly answers from memory. For “best CRM for dental practices 2026”, it usually searches. That is why a brand can be invisible in casual chats and still show up when buyers compare options, or the other way round.

Turning search on or off changes the answer, but by how much is disputed. In one 20k-response study, only 19.8% of products recommended with search off also appeared with search on (180 Marketing, ChatGPT search vs no search). Another study found little change. The methods differ (products vs brands), so the honest reading is that you have to test your own category.

How AI models know a brand: frequency, co-occurrence and cutoffs

A model knows a brand roughly in proportion to how often the brand appears in its filtered training text, especially next to its category and need words. The mechanism is strong evidence: question-answering accuracy rises with the number of training documents that contain an entity, and removing those documents hurts recall (Kandpal et al., ICML 2023). Entity popularity, measured by Wikipedia page views, predicts recall too (Mallen et al., PopQA, ACL 2023).

Three details shape what that means for a brand.

• Co-occurrence beats raw frequency. Being written about together with the category phrase (“dental scheduling software”) matters more than being mentioned at all. That is an inference from the research, graded plausible for brands.
• Filters decide what counts. Training pipelines drop thin, duplicate, JavaScript-only and boilerplate pages, and Common Crawl gives poorly linked domains a small crawl budget (confirmed from the published crawl and filter methods). Syndicated press releases are mostly deduplicated away.
• Memory is frozen per release. A model's knowledge changes only when a new version ships, and the last months before the cutoff are thinly represented (Cheng et al., COLM 2024). Release lag plus time in production means mentions earned today mostly reach memory 6–24 months later. That timeline is inferred from published cutoff dates, so plausible, not measured.

Thin training presence costs more than being left out. Code models invent package names in 19.7% of samples, and the same fake names repeat (Spracklen et al., USENIX Security 2025, 2.23M samples). A product the model barely knows gets replaced or made up.

How AI engines find a brand: fan-out queries and different search backends

An AI engine finds a brand by running its own rewritten searches, called fan-out queries, against a search backend, then reading short snippets from the results. The user's prompt is rarely the query. Fan-out queries are longer and often add “2026”, “price”, “reviews”, “vs”, “official” or a site: filter (strong evidence from logged queries). The glossary entry for fan-out explains the term.

The search backend differs by engine, so ranking well in Google covers only part of the picture.

Ranking for the parent prompt predicts little. Only about 8–12% of ChatGPT citations overlap Google's top 10 (Ahrefs, 15k prompts, Jul 2025), and for AI Overviews the figure ranges from 17% (BrightEdge) to 38% (Ahrefs, 863k SERPs, Mar 2026). Being indexed and ranking somewhere for the sub-queries matters more. Engines also mostly see snippets, not pages: free ChatGPT keeps roughly the first 202 characters of a page's rendered body (strong, from repeated fetch tests). optmz shows the fan-out searches each engine ran for every tracked question, which is where found problems usually become visible.

How to tell which problem you have: three known vs found tests

You can tell a known problem from a found problem by asking the same buyer questions with search off and with search on, then checking what the model believes about you. Run each test many times, because a single answer is noise (see how to measure AI visibility without fooling yourself).

1. Search off vs search on. Ask your buyer questions through a model with no web tools, then through the search-enabled surface. Count how often the brand appears in each, over at least 20 runs per question. Present with search off but missing with search on points to a findability problem. Missing with search off but present with search on points to a knowledge problem.
2. Need-specific prompts. Category-only prompts (“best CRM”) often leave out even large brands, and some brands appear only when the prompt adds need cues (“CRM for a three-chair dental practice that handles insurance claims”). A 2026 preprint across six models calls these brands conditionally retrievable (Malthouse et al., abstract read, plausible). If you only appear on narrow prompts, the model knows you, but weakly.
3. Entity fact check. With search off, ask for your founding year, headquarters, pricing tier, main product and competitors. Score the answers against a fact sheet. Wrong facts show what the model learned, and they point to the pages to correct, such as Wikidata, your about page or a review profile. This follows the PopQA approach of testing what a model knows about an entity.

optmz is building a Known vs Found check that runs this split for each tracked question. It is in development, so for now the tests above are manual.

What feeds known vs found: the brand footprint list

Each off-site source feeds training (T), live retrieval (R) or both, and the evidence behind each differs. The table rates each channel from the research, with the best evidence grade we found. The ratings are judgments, not measured coefficients.

Backlinks matter mostly indirectly. Common Crawl sizes each domain's crawl by a link-graph measure, so a rarely linked brand is sampled thinly before any filter runs (confirmed). Direct correlations between domain rating and AI visibility are weak (Ahrefs, ρ 0.22–0.33).

Blocking crawlers splits the two channels too. OpenAI documents GPTBot for training and OAI-SearchBot for ChatGPT search. Blocking GPTBot keeps your own text out of future training but does not remove you from ChatGPT search, and it does nothing about what others write about you (confirmed, OpenAI's crawler documentation).

What not to do: Wikipedia pages, fake reviews and self-promotional listicles

The common shortcuts for AI visibility either break platform rules or fade. Three come up most often.

• Writing your own Wikipedia page. Wikipedia's notability rules for companies forbid promotional page creation, and no causal study shows that creating a page raises recommendation rates. Claims like “50% more likely” come from vendors and anecdotes. Correcting sourced facts on Wikidata is allowed and low risk.
• Fake or incentivized reviews. They break review platforms' terms and risk removal, and review counts have a weak link to citations anyway (G2's own analysis of Profound data: about 1% of variance explained).
• Mass self-promotional listicles. Ahrefs ran a controlled experiment on its own brands (34 pages, 9,886 answers). The lists helped a new brand get mentioned, barely moved an established one, and the citations faded over time. In 43% of answers citing those pages, the brand was not mentioned at all. It is one company, so plausible, but it is the most honest test published.

What does work is slower and more ordinary: earned mentions in places your buyers read, honest and disclosed answers on Reddit and Stack Overflow, YouTube reviews and tutorials, and inclusion in third-party comparisons.

Known vs found for dev tools: usage in code counts more than marketing

For developer tools, the strongest known signal is usage in code, not marketing pages. A model trained on code sees a library every time a repository imports it, a Stack Overflow answer uses it or a tutorial walks through it.

Models favor established libraries. In one study, Polars was never chosen over Pandas, FastAPI appeared in 9% of responses, and these preferences barely moved across prompt variants (Twist et al., ACL Findings 2026, strong). An independent study of 2,430 coding-agent runs found default picks shifted sharply between model versions, with one UI library falling from 90% to 52% of picks (Amplifying, 2026, plausible). Three practical points follow.

• Test with code tasks (“write a script that does X”) and check the imports, not only “what is the best X”.
• Make the tool usable by agents: signup, API keys and docs an agent can follow. In the same study, one model wrote its own code instead of using a database service that needed keys only a human could create.
• Track dependents and downloads as a rough proxy for training presence. No causal study links GitHub stars or awesome-lists to recommendations.

Known vs found checklist for agencies

A known vs found review takes a few hours per client and decides which kind of work to sell. Use this order.

1. Pick 10–20 buyer questions, including need-specific versions.
2. Run each with search off and search on, at least 20 times per engine, and record mention rates with a confidence range.
3. Run the entity fact check with search off and list every wrong fact.
4. For found problems, check crawler access, server-rendered HTML and indexing in Google and Bing, then compare your pages with the fan-out searches engines run.
5. For known problems, map the footprint table to the client's vertical and plan off-site work in the channels that feed training.
6. Set expectations: retrieval fixes can show in weeks, while training effects arrive with model releases, often 6–24 months out.
7. Note the model version next to every search-off result, and rerun the panel when a new model ships.

For the retrieval side in detail, read what actually moves AI citations in 2026.