How to Get Your Business Recommended by ChatGPT

Two engines produce one answer

Ask ChatGPT “best payroll software for a 12-person agency” and one of two things happens. Either the model answers from its parameters — patterns compressed from training data that stopped at a fixed cutoff, documented per model in OpenAI's model reference — or it decides the question needs fresh information and calls its search tool, which retrieves live pages and returns an answer with links, as described in OpenAI's ChatGPT search documentation.

The consequences differ completely. If your company launched, rebranded, repriced or repositioned after the cutoff, the base model has no reliable representation of you and will happily recommend competitors it does remember. Nothing you publish today changes that until a future training run absorbs it. The search path, by contrast, is live: a page indexed this month can be cited tomorrow.

So the near-term lever is retrieval, and the long-term lever is being written about widely enough, on durable third-party sources, that the next model generation inherits you as a known entity. Both matter. Only one is fast.

Scale is the reason this is worth staffing. Pew Research Center found in a June 2025 survey that 34% of U.S. adults had used ChatGPT, roughly double the 2023 share — a meaningful slice of the population now treats it as a first-stop shortlist generator.

What ChatGPT search actually pulls from

OpenAI publishes three distinct crawlers in its bots documentation: GPTBot, which collects data that may be used for training; OAI-SearchBot, which builds the index used to surface and link to sites in ChatGPT search; and ChatGPT-User, which fetches a page when a user or an action triggers it. These are separate user agents with separate purposes. A lot of publishers blanket-blocked anything OpenAI-shaped in 2023 to opt out of training and, in doing so, opted out of being recommended.

OpenAI does not publish a ranked list of the sources its search layer prefers, and anyone who tells you they have the algorithm is guessing. What is observable from the answers themselves is the shape of the pages that get cited: they tend to be pages that already rank in conventional web search and that match the comparative form of the question. A “best tools for X” prompt retrieves “best tools for X” pages — editorial roundups, category directories, review platforms, forum threads — far more often than any vendor's own homepage.

Two practical implications follow. First, classic indexability hygiene still governs whether you're retrievable at all: server-rendered content rather than client-side-only rendering, no interstitials over the substance, clean canonicals. Keeping your site healthy in Bing Webmaster Tools and pushing changes through IndexNow costs nothing and improves your odds across every retrieval-based assistant. Second, machine-readable entity data helps: a schema.org Organization or LocalBusiness block that states your name, category, service area and identifiers gives a parser facts instead of prose to interpret.

Order of operations when ChatGPT has never heard of you

Do these in sequence. Steps three and four are wasted effort if step one is broken, and most teams start at four.

1. Confirm you are fetchable. Read your own robots.txt line by line for OAI-SearchBot and ChatGPT-User, then check your CDN or WAF bot rules — managed bot protection blocks these crawlers by default at several providers, and no one on the marketing team is told.

2. Fix the entity record before adding content. One canonical company name, spelled the same everywhere. One sentence on your homepage and About page that says what you sell, who it's for, and where you operate, in plain declarative language a model can lift verbatim. Consistent details across your site, your profiles and any directory that already lists you.

3. Get onto the pages that already answer the question. Run the ten prompts your buyers actually type, note which third-party pages ChatGPT cites, and pursue inclusion on those specific pages with verifiable facts — a spec sheet, pricing, a screenshot, a customer reference. This is outreach work, not content work.

4. Publish the comparative material your competitors avoid: real pricing, an honest “who this isn't for”, named alternatives, integration and limitation lists. Comparative prompts need comparative source text, and if you don't supply yours, the model uses somebody else's characterisation of you.

5. Build corroboration. Reviews with substance, forum answers written by a named human who discloses affiliation, mentions in trade press. Repetition across independent domains is what turns a claim into something an assistant will assert without hedging.

6. Re-test on a schedule and treat the source links in each answer as your backlog — every cited page you don't appear on is a concrete, nameable target.

Why third-party pages beat your own site in the answer

A recommendation is a judgement, and models are conservative about sourcing judgements to the party being judged. Your homepage is excellent evidence that you exist, that you serve mid-market logistics firms, and that you have an API. It is weak evidence that you are the best option, and ChatGPT's search layer behaves accordingly — it will cite you for attributes and cite others for verdicts.

That's why the highest-leverage work is usually off your domain. Being the seventh entry in a roundup that ChatGPT already cites for your category outperforms a tenth blog post on your own site, because it inserts you into the exact document the retrieval step is fetching. The same logic applies to review platforms and to communities where practitioners answer each other; those threads are dense with the comparative language these prompts match.

One caution: don't try to manufacture consensus. Seeded fake reviews and near-identical guest posts across low-quality domains create a fragile footprint, and if the corroborating pages disagree with your own site, assistants hedge or omit you rather than pick a side. Consistency of facts across independent sources is the asset.

Measuring whether any of it worked

ChatGPT answers are non-deterministic. The same prompt on the same day can produce different shortlists, so a single screenshot is anecdote, not measurement. What works is a fixed prompt set — 30 to 60 questions phrased the way your buyers phrase them, including the unbranded category questions where you're most likely absent — run repeatedly and scored on four things: whether you're mentioned, where in the list, how you're characterised, and which domains the answer cited.

The cited domains are the most actionable output. They convert a vague ambition into a target list of pages to appear on, and they show you when a single directory or review site is quietly acting as the gatekeeper for your entire category. Track the same set across assistants too, because a page that carries you in ChatGPT often carries you in Perplexity and Copilot as well — the broader picture is in our write-up on AI visibility.

This is exactly what we built Growli to do: run your prompt set across ChatGPT and the other major assistants on a schedule, score mention rate and sentiment, extract the cited sources, and hand back a ranked list of fixes. How it works walks through the mechanics if you'd rather not maintain the tracking spreadsheet yourself.

What not to spend the budget on

Skip anything sold as a guaranteed placement inside ChatGPT recommendations. There is no ad slot in the organic answer and no submission form; anyone promising a listing is selling you a directory entry with a markup.

Be sceptical about llms.txt as a growth tactic. It's a reasonable convention and harmless to publish, but OpenAI's own crawler documentation doesn't mention it, so treat it as optional plumbing rather than a channel. The same goes for hidden instructions aimed at models — text a human can't see is a policy problem, not a strategy, and it can get a domain distrusted.

Finally, resist mass-producing thin pages. Retrieval rewards being the best source for a specific comparative question, not having the most URLs. Ten pages that state prices, tradeoffs and named alternatives will out-earn two hundred that restate your value proposition.

FAQ