This is the most common question we get, and most answers to it are either vague or sell you a monthly subscription. So here is the plain version, based on what actually correlates with getting named in AI answers.
Start with a reframe. There is no position one to win. When we asked four engines to recommend products across eight categories, they agreed on a single best pick in only one of them. So the goal is not to rank. The goal is to be the answer the models keep arriving at independently, from different sources, often enough that you are hard to leave out.
This is not theory dressed as a playbook. Our open research includes the methods and downloadable summaries behind the 96-answer recommendation study and the 180-answer household-brand audit.
What actually drives a recommendation
AI engines build their answers from the open web, third-party sources, and the communities people trust. The single strongest signal in the public research is not your own site. It is how much the rest of the web talks about you.
- ✓Third-party mentions and reviews. Listings on the sites buyers already check (G2, Capterra, industry roundups) are raw material the models read and repeat.
- ✓Comparison and alternatives content. The best X for Y and alternatives to Z pages are exactly the queries buyers ask AI, so being present in that genre matters a lot.
- ✓Communities the models cite. Reddit is now among the most-cited domains in ChatGPT and Perplexity answers. Honest presence there is read back to buyers.
- ✓Your own content, made quotable. Clear claims, real numbers, and direct answers to buyer questions give the model something clean to lift.
- ✓Accessible, unambiguous pages. Put important facts and evidence in visible text, keep internal links strong, and remove unintended search-access barriers. Permission alone does not guarantee selection.
The levers, in priority order
- Freeze a buyer-question baseline so you know which recommendation, citation, competitor, or accuracy gap you are solving.
- Fix unintended search-access, rendering, indexability, internal-link, and factual-consistency problems. Treat training access as a separate policy choice.
- Earn presence in the third-party sources and comparison content for your category. This is the heaviest lever and the slowest, so start now.
- Show up authentically in the communities your buyers use. Helpful, not spammy, because the manipulative version is getting caught and penalized.
- Make your own pages the cleanest source on the questions you want to win, with specific, quotable answers near the top.
- Then run a matched measurement across engines, because one memorable screenshot is not a baseline.
Remember the difference between mentioned and recommended. In our test, HubSpot was named in almost every small-business CRM answer and was still the top pick for none of the four engines. Getting listed is step one. Owning the first slot takes a concentration of strong, consistent signals that most competitors never build.
What does not move the needle much
A few popular tactics are mostly noise. Keyword stuffing your pages does little, because these models read for meaning, not density. Shipping an llms.txt file will not lift you on its own, whatever a generator promises. And a single PR hit fades fast unless it feeds durable third-party presence. Spend your effort on the levers above instead.
A 30-day starting plan
- Week one: map the buyer questions, capture the baseline, and fix unintended search-access, rendering, or factual conflicts.
- Week two: claim and complete every relevant third-party listing and review profile, and make sure your details are current and consistent across them.
- Week three: improve the two decision pages most tied to your measured prompt and citation gaps, using current primary sources and honest tradeoffs.
- Week four: test. Run buyer-intent prompts across ChatGPT, Claude, Gemini, and Perplexity, a few times each, and write down where you appear and who beats you.
The free prompt pack generator builds that week-four test for you, grouped by buyer intent. And when you want the gaps found and ranked instead of eyeballed, that is the audit: the same prompts across three engines, scored, with the fixes ordered by impact.