Guide

How to get ChatGPT to recommend your business

Published July 23, 2026 · Opinion Radar

ChatGPT and its peers already hand out recommendations thousands of times a day — CRMs, agencies, restaurants, databases. Those recommendations are measurable, surprisingly concentrated, and they can be influenced. Not with tricks: with the same signals the models actually read. Here is what our measurements show, and the eight levers that work.

How ChatGPT decides what to recommend

Two mechanisms produce a recommendation, and they respond to different levers:

• Training memory. The model has read the web up to a cutoff date and retained associations ("CRM for a small team → these tools"). This moves slowly — with model updates — and reflects your accumulated reputation: comparison articles, reviews, community threads, press.
• Live web search. For fresh or factual questions, the assistant searches the web and cites sources. Here everything depends on which pages it finds at answer time, and how easily facts can be lifted from them. ChatGPT's search draws on third-party search providers, including Microsoft Bing — a site invisible in classic search engines is invisible to it too.

Both mechanisms share one property: the answer is a consensus of third-party sources, not the brand's own messaging. That single fact explains most of what works and what doesn't.

What the measurements show

Three practical conclusions. AI recommendations are concentrated (a few names take almost everything), consensual (models largely agree), and earned off-site (the sources that drive them are third-party pages, not your homepage). See it live in our database observatory — PostgreSQL takes 58% of 430 answers — or our study of whether AI assistants recommend themselves.

Independent citation studies point the same way. Profound's analysis of 680 million AI citations (2024-2025) found Wikipedia to be ChatGPT's single most-cited domain (7.8% of citations), with review platforms like G2 in the top ranks — encyclopedias, communities and review sites, not brand homepages. And Peec AI's analysis of 30 million sources (March 2026) puts Reddit, YouTube and LinkedIn at the top across AI platforms.

The eight levers that actually work

1. Let AI crawlers in. Check your robots.txt for GPTBot and OAI-SearchBot (OpenAI documents them separately — allowing one does not allow the other), ClaudeBot (Anthropic), PerplexityBot and Google-Extended (which, per Google's docs, does not affect your Google Search rankings). Blanket bot-blocking — often inherited from an old template — removes you from the pool of citable sources.
2. Put answers in plain HTML. AI crawlers don't execute JavaScript — Vercel's analysis of AI crawler traffic on its network concluded that "none of the major AI crawlers currently render JavaScript" (December 2024). If your content only exists after a client-side render, it may be a blank page to them. Server-render what matters.
3. Build one page per buying question. People ask assistants situations, not keywords: "best CRM for a 5-person agency". A page that squarely answers one situation is the natural source for that answer.
4. Publish quotable, dated facts. Generative models prefer precise, sourced claims ("X% in 2026, according to…") over slogans. Pricing, limits, benchmarks, dates: give the model something liftable.
5. Get into the comparisons AIs actually cite. "Best X in 2026" listicles and independent comparison sites dominate the citations we measure. Being absent from them means being absent from the models' raw material. This is outreach work, not on-site work.
6. Cultivate reviews and communities. G2, Capterra, Trustpilot — and Reddit, which shows up relentlessly in cited sources. The good news: you don't need viral threads. Semrush's study of 248,000 Reddit posts cited by AI found that 80% of cited posts have fewer than 20 upvotes, and the average cited post is about 900 days old — old, substantive Q&A threads are what gets cited. You don't "post ads" there; you earn genuine user mentions, and you answer honestly where your customers already talk.
7. Keep your entity consistent. Same name, same description, same facts across your site, directories, social profiles, Wikipedia/Wikidata where eligible. Models consolidate identity poorly when the record is inconsistent.
8. Show freshness. Visible, honest update dates. Answer engines discount stale sources — and buying questions are almost always asked in the present tense.

What doesn't work (despite what you've been sold)

• llms.txt. The server logs are in: Ahrefs analyzed 137,000 sites with an llms.txt file and found that 97% of those files received zero traffic — nothing fetches them (June 2026). Google's John Mueller has been saying it plainly: "no AI system currently uses llms.txt". Meanwhile adoption has grown almost 9× in a year — a textbook cargo cult. Harmless at best.
• "Magic schema for AI". Structured data helps search engines (a fine reason to do it well); there is no secret markup that makes a model recommend you.
• Keyword stuffing, AI edition. Repeating "best X" across your pages does not influence a generative synthesis — and it degrades your regular SEO.
• Hidden text aimed at models. Invisible instructions or white-on-white text: ineffective, and indistinguishable from cloaking, with the risks that carries.
• Mass-rewriting your site "for AI". There is no measured basis for a wholesale "AI style" rewrite. The content levers above are targeted, not cosmetic.

Measure before you optimize

You cannot manage what you cannot see. Before investing anywhere, establish the baseline: do assistants recommend you today? On which questions? With which arguments, which factual errors, and who do they name instead? A serious measurement needs three things: questions that never prompt your brand name (spontaneous mentions only), several model families (they genuinely disagree), and repeated sampling (answers vary run to run — one screenshot is not a measurement).

And measure it over time, not once: AI answers are a moving target. Semrush's 13-week study of 230,000 prompts watched Reddit's share of ChatGPT citations collapse from roughly 60% of answers to about 10% in six weeks — the sources behind AI answers get reshuffled faster than any search algorithm update.

That is exactly the protocol behind Opinion Radar, documented on our method page and applied in public in our observatories: every figure drills down to the raw answers behind it.

FAQ

Can you pay ChatGPT to recommend your business?

No. There is no advertising program that buys recommendations inside assistant answers. Recommendations are synthesized from the model's training data and from the sources it reads at answer time — which is why earning presence in those sources is the actual lever.

How long does it take to appear in ChatGPT answers?

Web-grounded answers can pick up new sources within weeks, because the assistant searches the live web. Training-data effects are slower — months, tied to model updates. Measure monthly and expect gradual movement, not a switch.

Does llms.txt help you get recommended?

No. Ahrefs' server-log study of 137,000 sites (June 2026) found that 97% of llms.txt files receive zero traffic — nothing fetches them — and Google's John Mueller has stated that no AI system currently uses llms.txt. Focus on being crawlable, quotable and present in the third-party sources assistants actually cite.

How do I know if ChatGPT recommends my competitors?

Ask the questions your buyers ask — without naming any brand — several times per model, and count who gets named. That is exactly what an AI visibility measurement automates across models, questions and repetitions.