How long until a fix shows up in answers?

It depends which cause you fixed, and that is this list's most useful practical consequence. Retrieval-side causes (3 and 9) are measurable in days to weeks, because once the page is re-crawled the next answer works from the new state. Causes rooted in parametric memory (2 and 4) refresh only with a new model generation — months is the unit there, and your work lands in the next training round rather than in today's answer.

I asked once and I was there. Am I fine?

One run is not a measurement. The same question can return a different answer in the same hour, because these models are not deterministic. To know where you stand, ask the same question repeatedly and read the result as a proportion rather than a yes/no — and vary the order of your questions, since an earlier question can influence the next one.

Isn't good content enough?

Necessary but not sufficient, and the order matters. If a machine cannot bind your name to a single thing (cause 1), even excellent content attaches to an ambiguous entity — which is to say, not to you. Entity work and access are the two foundations content cannot work without. Past those, content quality genuinely is the differentiator.

My competitors appear and I don't. Does that tell me anything?

Yes — it is one of the cheapest diagnostic signals available. If competitors come back with citations, the question is answerable and the retrieval layer is working, so the gap is on your side rather than in the question. Look at which of their pages the answer cites: that shows you what format and depth that particular surface finds quotable.

What if my content gets cited but my name is never written?

That is its own failure class, and more common than people expect — it happened to us: one surface cited our own research in five runs out of five while our company name appeared in the answer not once. The usual cause is that the content never says in its VISIBLE text who produced it. An author field in the structured data is not enough; the name and the role have to be in the sentences a reader sees.

Cookbook

Why doesn't ChatGPT recommend your brand? Nine causes, in order

Nine ordinary reasons an assistant leaves you out — and the one most teams fix first is usually the wrong one.

Updated · August 9, 2026

The question almost always arrives as "what should we write so AI recommends us?", and the ordering is almost always wrong. An assistant answers from two independent sources: what the model KNOWS about you from training, and what it FINDS about you at the moment of answering. These not only can diverge, they routinely do. Our own measurement in August 2026 came out like this: across forty closed-book questions the brand name appeared zero times, while on a sourced surface the same week our own research was cited in five runs out of five — without our name next to it. Same brand, same week, two opposite diagnoses. So this list is built for elimination rather than inspiration. Work top to bottom; each cause has one check that tells you whether it applies to you, and one fix that actually addresses that cause. Do not jump to the ninth until something measured has ruled out the first.

The procedure

Entity ambiguity — the machine cannot tell which one you are
The most common cause, and almost never the one people look for. If your name also fits other real things — another company, a city, a product, a common word — the signal about you scatters across your namesakes and the model commits to none of them. CHECK: ask for your name alone, with no context, on several surfaces, and see what it guesses. Mixed results, or a confident description of someone else, is your answer. FIX: one canonical name plus disambiguators (domain, city, category) emitted consistently from a single source — the entity canonicalization procedure.
Corpus absence — there is nothing to know you from
A model learns from text that was public when it was trained. If effectively only your own website describes you, a closed-book question has nothing to recall. CHECK: count how many INDEPENDENT domains make a substantive statement about you — not a mention, a claim about what you do. If that number is one or two, this cause is live. FIX: presence in independent sources where your name carries a factual assertion. Independence matters more than volume: a third article from the same publisher adds far less than a first mention on a new credible domain.
Retrieval absence — nothing finds you at answer time
The inverse of the previous cause, and it must be measured separately. A model that does not know you from memory could still cite you, if you were findable and quotable in the retrieval layer. CHECK: ask the same question on a surface that shows its sources, and read the source list. Competitors present and you absent means the problem is retrieval, not training. FIX: quotable claim density — specific, self-contained, sourced sentences rather than general marketing prose — plus crawlable, server-rendered pages.
Stale knowledge cutoff — the model knows an older you
If your name, positioning or focus changed in the last year or two, the model will happily describe the old one. That is not a defect; it is the timestamp of its training. CHECK: ask closed-book what your company does and compare with today. A description of a previous incarnation identifies this cause. FIX: consistency and patience. Parametric memory refreshes only with a new model generation; until then, strengthen the retrieval side so at least the sourced answer is current. What you cannot do here is force forgetting.
Contradictory claims — you assert several truths at once
When your homepage, your professional profiles, a company registry and your press kit each state a different founding year, headquarters, team size or category, the model cannot adjudicate, and it often resolves that uncertainty by staying silent about you. CHECK: put your five core facts side by side from every surface that states them. FIX: one source they all derive from, with the others pointing back to it. Contradiction costs more than omission: a missing fact is skipped, while a contradicted one is treated as uncertainty about the whole entity.
Language gap — you exist in one language and not the other
Answers in different languages are built from different corpora and frequently different sources. A brand present only in its local language does not surface for English questions, and the reverse holds too. CHECK: run the same buyer question in both languages and compare who appears. FIX: in the missing language, write rather than translate — the target questions genuinely differ, and machine translation carries the source language's search habits into a market that does not share them. The gap is also an opportunity: competition there is usually far thinner.
Category misplacement — you are filed where nobody asks
Sometimes the model knows you, but under a different category — so you are absent from exactly the questions where you would compete. CHECK: ask what your company does and which category it belongs to; then ask about that category and see whether you come back. FIX: the category word has to appear where machines read — in visible text, in the opening paragraph, consistently across surfaces. Classification does not improve because you know it; it improves because it is written down.
No evidence — there is nothing to justify recommending you
A recommendation needs a reason. If all that can be known is that you exist, but not what you worked on, for whom, or to what end, the answer's author picks whoever there is something to say about. CHECK: read the sentences about you and ask whether they contain a checkable fact or only adjectives. FIX: specific, dated, checkable statements — what you did, when, in what setting. Reviews and third-party assessments count here; an unsourced superlative on your own site does not, because there is nothing for a machine to attach it to.
Technical block — they never got in
Last on the list because it is the rarest — but when it applies, every fix above it is inert. CHECK: evaluate your robots.txt against actual AI user-agents with a parser rather than by eye; confirm the content renders server-side; and look in your server logs to see whether those bots arrive at all. FIX: sort out access, but mind the resolution — training, retrieval and user-initiated bots are separate classes, and blocking training does not remove you from search. There is a separate recipe for exactly this.

Frequently asked questions

How long until a fix shows up in answers?: It depends which cause you fixed, and that is this list's most useful practical consequence. Retrieval-side causes (3 and 9) are measurable in days to weeks, because once the page is re-crawled the next answer works from the new state. Causes rooted in parametric memory (2 and 4) refresh only with a new model generation — months is the unit there, and your work lands in the next training round rather than in today's answer.
I asked once and I was there. Am I fine?: One run is not a measurement. The same question can return a different answer in the same hour, because these models are not deterministic. To know where you stand, ask the same question repeatedly and read the result as a proportion rather than a yes/no — and vary the order of your questions, since an earlier question can influence the next one.
Isn't good content enough?: Necessary but not sufficient, and the order matters. If a machine cannot bind your name to a single thing (cause 1), even excellent content attaches to an ambiguous entity — which is to say, not to you. Entity work and access are the two foundations content cannot work without. Past those, content quality genuinely is the differentiator.
My competitors appear and I don't. Does that tell me anything?: Yes — it is one of the cheapest diagnostic signals available. If competitors come back with citations, the question is answerable and the retrieval layer is working, so the gap is on your side rather than in the question. Look at which of their pages the answer cites: that shows you what format and depth that particular surface finds quotable.
What if my content gets cited but my name is never written?: That is its own failure class, and more common than people expect — it happened to us: one surface cited our own research in five runs out of five while our company name appeared in the answer not once. The usual cause is that the content never says in its VISIBLE text who produced it. An author field in the structured data is not enough; the name and the role have to be in the sentences a reader sees.

Sources

OpenAI — Bots (GPTBot, OAI-SearchBot, ChatGPT-User). https://developers.openai.com/api/docs/bots
Google — Google crawlers (Google-Extended). https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers
Perplexity — Perplexity Crawlers. https://docs.perplexity.ai/docs/resources/perplexity-crawlers

Why doesn't ChatGPT recommend your brand? Nine causes, in order

The procedure

Entity ambiguity — the machine cannot tell which one you are

Corpus absence — there is nothing to know you from

Retrieval absence — nothing finds you at answer time

Stale knowledge cutoff — the model knows an older you

Contradictory claims — you assert several truths at once

Language gap — you exist in one language and not the other

Category misplacement — you are filed where nobody asks

No evidence — there is nothing to justify recommending you

Technical block — they never got in

Frequently asked questions

Sources