How to get ChatGPT to recommend your brand

These are practical suggestions, not disclosed ChatGPT ranking factors. Google's people-first content guidance supports clear, useful, attributable information in Search; it does not establish a ranking formula for ChatGPT.

First, the honest caveat: nobody can guarantee ChatGPT will recommend you. Answers vary between runs, models update, and the ranking logic is not public. What you can do is stack the observable odds in your favour. Everything below is based on how these systems visibly behave, not on secret tricks.

1. Find out where you stand today

For an unbranded discovery baseline, use buyer questions such as "best [category] for [audience]" and "which tools support [need]?" without naming your brand or a competitor. Record mentions, position, citations, model and sample count. Arcanoris asks each question once per provider per audit; repeated comparable audits provide additional observations. See our measurement methodology.

Write the questions from the buyer's vocabulary, not yours. "Best CRM for a five-person agency" is what someone actually types; "enterprise-grade customer relationship management platform" is what your homepage says. If your prompt list reads like your own marketing copy, you're measuring how well you describe yourself, not whether buyers find you.

Branded comparison research is a separate exercise: prompts such as "[your brand] vs [competitor]" or "alternatives to [competitor]" can reveal perceived differences. Because they supply brand names, do not mix those results into an unbranded discovery mention rate.

2. Win the sources the answers already cite

When ChatGPT searches the web before answering, it leans on a small set of pages per question: comparison posts, review aggregators, community threads, industry lists. Those citations are a literal to-do list. Get reviewed on the aggregators that keep appearing. Pitch the authors of the comparison posts that omit you. Answer the recurring questions in the communities that get cited. This is classic digital PR, aimed with unusual precision.

The precision is the point: instead of guessing which twenty publications matter for your category, your own sampled answers tell you exactly which five to seven pages keep coming up. Rank them by how often they're cited across your question set and work down the list in that order - that's a better prioritization signal than any generic "top publications in your industry" list a PR agency will sell you.

3. Make your own site quotable

• State plainly what you are, who you're for, and what you cost. Models synthesize; give them clean sentences to synthesize from.
• Publish an honest comparison page. If you don't describe how you differ from competitors, a third party will do it for you, less favourably.
• Keep pricing public and current. "Contact us" pages give an assistant nothing to say when someone asks about cost.
• Use accurate structured data where it describes visible content. It does not guarantee that an assistant will use or cite the page.

A minimal FAQ schema block is a few minutes of work and directly answers the shape of question buyers ask an assistant:

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "How much does [Product] cost?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "[Product] starts at $29/month for up to 5 users, billed monthly or annually."
    }
  }]
}

4. Don't block the crawlers

Make separate choices about training, search crawling and user-requested retrieval. ClaudeBot is not Claude-SearchBot or Claude-User, and Google-Extended is not the control for AI Overviews. Server-rendered facts are easier to inspect, but do not assume every crawler has the same JavaScript capabilities. Our crawler guide explains the controls and links to provider documentation.

Check both robots.txt and your CDN or firewall logs. An allowed path can still be blocked at the network edge. The crawler guide includes a diagnostic request and explains why a spoofed user agent is not proof that a verified provider bot can fetch a page.

5. Feed the durable record

Keep accurate, attributable information on relevant third-party sites as well as your own. We cannot know whether a particular page will enter a future training set or how much it will affect an answer. Treat this as documentation and discoverability work, not a promise of persistent model memory.

6. Re-measure on a schedule

Repeat the same question set with the same providers and method. A single scan is a snapshot; a series can reveal changes but does not establish that your edits caused them. Model updates, retrieval changes and sampling variation are alternative explanations. Keep branded research separate from the discovery series.

Mistakes that quietly cancel all of this out

• Changing the question wording between measurements. A different prompt is a different experiment - you can't tell if your mention rate moved because of your work or because you asked a different question.
• Optimizing for one provider only. A page written purely to please one assistant's search behavior often does nothing for the others; track providers separately and expect the gains to arrive unevenly.
• Publishing a comparison page and never updating it. A stale comparison page is worse than none - it gets cited with numbers that are now wrong, and you don't control the correction.
• Treating one good run as proof it worked. Wait for the next full sample before declaring victory; a single favorable answer is exactly the noise this whole approach exists to filter out.