On this page
You ask ChatGPT “what’s the best tool for X” where X is exactly what your product does. It lists five tools. Yours is not one of them. The obvious question is: how did it pick those five? The answer is not “there’s a ranking and you’re on page two.” There is no ranking. Understanding what actually happens when a model generates a recommendation is the difference between guessing at AEO tactics and doing the few things that work.
There is no list inside the model
The first thing to unlearn: an LLM does not look up “best CRM for startups” in a database and return the top entries. It generates an answer one token at a time, and brand names come out because they are statistically strong completions in that context.
During training, the model consumed a huge slice of the public web: blog posts, documentation, Reddit threads, review sites, comparison articles, news. Every time your brand appeared near words like “project management,” “affordable,” or “for small teams,” the association between your name and that context got a little stronger. Ask the model for a project management tool later, and the brands with the strongest accumulated associations are the ones most likely to be generated.
This has two blunt consequences for a small SaaS:
- Frequency matters. A brand mentioned in 400 independent places outweighs a brand mentioned in 12, almost regardless of how good either product is.
- Context matters as much as frequency. If most mentions of your product sit next to the wrong descriptors, the model learns the wrong association. A time-tracking tool that mostly gets discussed in freelancer forums will surface for “time tracking for freelancers” and stay invisible for “time tracking for agencies,” even if agencies are your core market.
If your product launched after the model’s training cutoff, or was simply too small to leave a footprint in the training data, the model may not know you exist at all. We covered that failure mode in detail in why ChatGPT doesn’t know about your SaaS.
Training data is only half the story: retrieval
Modern assistants rarely answer purchase-intent questions from memory alone. Perplexity always searches the web. ChatGPT and Gemini frequently do, especially for “best X” and “X vs Y” questions where fresh information matters. The model runs one or more searches, reads the top results, and writes its answer grounded in those pages. This is the RAG pattern: retrieval-augmented generation.
This changes the game in your favor, because retrieval is something you can influence on a timescale of weeks, not training cycles:
- The assistant searches - often rewriting the user’s question into several sub-queries behind the scenes.
- A conventional search index returns pages - so classic SEO still gates which sources the model even sees.
- The model reads and synthesizes - and here, extractable content wins. Pages with clear comparison tables, explicit “who is this for” statements, and concrete numbers get quoted; vague marketing copy gets skipped.
- Citations get attached - Perplexity always, ChatGPT and Gemini when browsing.
The practical takeaway: the pages that dominate retrieval for buyer questions are listicles and comparisons (“best X for Y,” “A vs B”). If none of those pages mention you, the model synthesizing from them will not mention you either. Publishing your own honest comparison and best-of pages, and getting into third-party ones, is the most direct retrieval lever available to a small team.
The signals that tip the decision
Put training and retrieval together and you can enumerate what actually moves a model toward recommending a specific brand. Roughly in order of leverage for an indie SaaS:
- Third-party mentions in category context. Review platforms, niche blogs, newsletters, Reddit and Hacker News threads, YouTube walkthroughs. Independent sources count for more than your own site, and the model has no way to know a “top 10” listicle was written by a competitor, so the ecosystem of pages around a category is the battlefield.
- Consistency of your entity. Models resolve brands as entities. If your homepage says “AI-powered workflow platform,” your G2 profile says “automation tool,” and your Crunchbase says “productivity app,” the associations get diluted across three fuzzy identities. One tight, repeated one-line description of what you are and who you serve strengthens a single entity instead.
- Extractable facts. “Starts at $19/month, unlimited seats, SOC 2 certified” survives synthesis. “Flexible pricing for teams of any size” does not. Models preferentially repeat specific, citable claims because those claims are what make an answer look substantive.
- Structured data and crawlability. Schema markup, clean semantic HTML, and not blocking AI crawlers in robots.txt determine whether your pages are readable evidence at all. This is table stakes rather than an edge, but plenty of sites fail it.
- Sentiment of the surrounding text. A brand that mostly appears in “why I switched away from…” threads develops negative associations. Models do pick this up, and it shows in hedged or qualified recommendations.
Notice what is not on the list: your Google rank for the head keyword, your ad spend, your Domain Rating in isolation. These correlate with visibility but none of them is the mechanism.
Why the answer changes every time you ask
Run the same prompt five times and you will get overlapping but different brand lists. This is not a bug you can fix; it is how generation works. Models sample from a probability distribution, so a brand with a moderate association shows up in some runs and not others. Retrieval adds more variance: different runs can pull different pages.
This matters for measurement. A founder who checks once, sees themselves mentioned, and concludes “we’re fine” is reading noise. The meaningful metric is mention rate across repeated samples: out of 20 runs of a prompt, how many include you? That is the logic behind multi-sampling, and it is why any serious visibility number is an average over many runs and several days, not a screenshot.
Model updates add a slower layer of churn on top. Providers ship new model versions and retrievers change ranking behavior, so a brand’s mention rate can shift noticeably in a week without the brand doing anything.
What you can actually do about it
Condensing the mechanics above into a working plan:
- Write down the 20-50 questions your buyers actually ask. Not keywords, questions: “best changelog tool for small SaaS teams,” “alternatives to X that support SSO.”
- Fix your entity. One canonical description, used verbatim on your homepage, review profiles, directories, and social bios.
- Make your site evidence-grade. Specific facts, comparison pages, FAQ content, schema markup, crawlable to AI bots.
- Earn third-party mentions in category roundups. Review platforms first, then niche blogs and communities where your category gets discussed.
- Measure on a schedule, not on a whim. Run your question set against ChatGPT, Claude, Perplexity, and Gemini repeatedly and track mention rate over time. You can do a first pass by hand in an afternoon following this guide to tracking your brand in ChatGPT, and it is worth doing once just to feel the variance yourself.
Manual checking stops scaling almost immediately, though: 30 prompts across 4 models with 3 samples each is 360 conversations per polling cycle. That is the boring, repetitive part that AskAiRank automates - it runs your prompts on a schedule across all four assistants, detects mentions, and shows your mention rate and position against competitors so you can see whether the work above is actually moving the needle.
The takeaway
LLMs recommend brands by combining learned associations from training data with whatever evidence retrieval hands them at answer time. Both layers are shaped by the same underlying asset: a web-wide trail of consistent, specific, third-party-corroborated statements about what your product is and who it is for. Build that trail deliberately, then measure mention rate over repeated samples instead of trusting single conversations. Start today with the manual version: ask each assistant your five most important buyer questions, note who gets mentioned, and you will know within an hour where you stand.
Frequently asked questions
Answers about the signals ChatGPT, Claude, and Perplexity weigh before naming a brand.
No. There is no index with ordered positions inside the model. A recommendation is generated token by token, based on how strongly your brand is associated with the category in training data, plus whatever retrieved web content the assistant pulls in at answer time. That is why the same question can produce a different list tomorrow.
No. There is currently no paid placement inside ChatGPT, Claude, Perplexity, or Gemini answers. The only lever you have is the evidence layer: what the model learned about you during training and what it can retrieve about you from the live web.
For retrieval-backed assistants like Perplexity and ChatGPT with browsing, changes can show up within days or weeks of your content being crawled. For the model's internal knowledge, you are waiting on a training cutoff, which can lag by many months. This is why retrieval-facing content is the faster lever for small brands.
Google rankings measure page relevance to a query. LLM recommendations lean on entity strength: how often and how consistently your brand appears next to the category across reviews, comparisons, forums, and docs. A competitor with more third-party mentions can beat you in AI answers while losing to you in organic search.
Not really. LLM outputs vary between runs, so a single ask can show you a list you would not get five minutes later. You need repeated sampling across multiple prompts and multiple models before the picture is trustworthy.
