How to get your startup recommended by ChatGPT, Perplexity and Google AI (2026)

Updated September 2026 · 8 min read

Assistants recommend products the way a careful colleague would: they retrieve pages they can read, weigh what those pages and other sites say, and name the options that answer the question asked. So a startup gets recommended by ChatGPT, Perplexity and Google's AI features by being readable to their crawlers, by publishing pages that answer buyers' actual questions completely, and by being described consistently across the web. Google's official 2026 guidance on generative AI features says the same thing in different words: optimising for AI search is SEO, and the fashionable tactics sold as AEO or GEO mostly do nothing. This guide is the practical version, with the numbers we could verify and the tricks you can skip. We build Clerion, an analytics product that measures the result, so weigh that where it is relevant.

How an assistant decides what to name

Two mechanisms matter, and Google has described both.

Retrieval. When you ask an assistant for a tool that does X, it runs searches, fetches the pages that come back, and writes an answer grounded in what it read. Google calls this retrieval-augmented generation and says its AI features draw on the normal search index and ranking systems. ChatGPT search and Perplexity work the same way with their own indexes. If your pages are not in the index, or are in it but unreadable, you are not in the answer.

Query fan-out. The assistant does not run one search. It breaks the question into several: "best X for Y", "X pricing", "X vs Z", "does X need a cookie banner". Each fetches pages. A startup with one homepage and nothing else appears for none of them. A startup with a pricing page, two comparison pages and an explainer can appear for four.

This is why the pages assistants send visitors to are rarely homepages. Look at the landing pages of your own AI referrals and you will usually find comparison pages, pricing and plain explainers, because those are the pages that answered the question inside the assistant.

Step 1: be readable

Nothing below matters if the crawlers cannot read you. Three checks, covered in full in how to make your website readable by AI crawlers:

• robots.txt allows OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User and PerplexityBot. Blocking GPTBot for training is a separate decision and does not cost you citations.
• No CDN or firewall rule is returning 403s or challenge pages to those crawlers.
• The text of every public page is in the HTML the server sends. A study by Vercel and MERJ of more than 500 million GPTBot fetches found no JavaScript execution; the same holds for the Anthropic and Perplexity crawlers. Googlebot is the exception.

The free AI crawler checker runs these checks and four more in about ten seconds.

Step 2: publish the pages the questions land on

Map the questions a buyer asks an assistant before choosing a tool like yours, then make one complete page for each. For most software startups the list is short.

Each page should answer completely on its own, because the assistant may read only that one page. A 2026 analysis of cited pages found that whether a page is a complete, self-contained answer predicted citation better than any other factor it measured. Put the answer in the first paragraph, then the detail, then a short FAQ.

Google's guidance is blunt about quality here. Content it calls commodity, the "7 tips for first-time buyers" kind that anyone could write, adds little. Non-commodity content brings something only you have: your own data, your own comparison with prices you checked, a first-hand account of a problem. Write the pages yourself from what you know. If you use AI to draft, the result still has to say something a model could not have said without you.

Step 3: be described consistently elsewhere

Assistants weigh what other sites say about you, and several 2026 studies found that mentions of a brand across the web correlate with being named in AI answers more strongly than most on-page factors. One analysis of citation sources across ChatGPT, Perplexity, Claude, Gemini and AI Overviews found Reddit, Wikipedia, YouTube and LinkedIn among the most cited domains by a wide margin.

The honest way to use this:

• Be present where your buyers already talk. Answer questions in the forums and communities where people ask about your category, as yourself, with your product named when it is relevant and other options named when they fit better.
• Get listed in the directories and comparison articles that already rank for your category's queries. Those pages are what the fan-out searches return.
• Say the same thing everywhere. Product name, one-line description, price. If your site says $14 a month and a directory says $19, the assistant picks one and it may not be yours.
• Fix your name if it collides. Clerion collides with Celerion, a clinical research company, and Clarion, a car audio brand. Assistants returned those instead of us. We published a plain page on what Clerion is and is not, added alternate names and a disambiguation to our Organization schema, and put the same paragraph at the top of our llms.txt. If your name is a common word or close to another company's, do this first.

Google's guidance names the opposite tactic, seeking inauthentic mentions, as something its spam systems handle. Bought mentions on sites nobody reads do not move the retrieval step.

What you can skip

Google published a short list in 2026 of tactics its AI features do not use. Nothing suggests the other assistants use them either.

• llms.txt. Google ignores it. An Ahrefs study of 137,000 domains found 97 percent of published files were never fetched. Harmless, occasionally useful for agents already on your site, not a visibility lever. See what is llms.txt and do you need one.
• Chunking. Splitting pages into fragments for models to digest. Models read long pages fine.
• Rewriting for AI. Question-shaped headings on every paragraph, keyword variants for every phrasing. Models understand synonyms. Google treats scaled variants as spam.
• Special schema. There is no AI-specific structured data. Ordinary schema that describes a real thing is still worth having for rich results and for getting facts right.
• Tools claiming internal Google metrics. No third party has them.

Step 4: measure what happened

Three signals, in the order they appear.

Crawls. The assistant crawlers fetching your pages, visible only in server logs or your CDN's bot analytics. This is the leading indicator: a page fetched this month can be cited next month.

Google's report. Search Console has a Generative AI performance report showing how your pages appear in AI Overviews and AI Mode. Your property has to be included in Search generative AI features in Search Console to be eligible; check that setting.

Referrals. People clicking from an answer to your site. ChatGPT tags links with utm_source=chatgpt.com; Perplexity, Claude and Gemini pass their domains as referrers when they pass anything. Similarweb's 2026 report put AI referral traffic at 770 million monthly visits worldwide, up 117 percent year over year. Still a small share of the web, but concentrated on exactly the kind of pages this guide asks you to publish. How to track ChatGPT traffic to your website covers the Google Analytics 4 setup. Clerion reports AI assistants as their own channel without setup and reads the channel alongside everything else, so you can ask which pages assistants send people to and whether those visitors convert; plans start at $14 a month with a 14-day trial.

Frequently asked questions

Is AEO or GEO different from SEO?

Not according to Google, whose AI features run on its search index and ranking systems. The other assistants also retrieve from indexes built by crawlers that honour robots.txt and read HTML. The work is the same: be crawlable, publish complete pages that answer real questions, be described consistently elsewhere.

How long does it take to get recommended?

There is no published timeline. In our experience the order is weeks for crawlers to fetch new pages, then months for those pages to appear in answers, and it depends on how many other sites already answer the same question well.

Should I write pages for every question variant?

No. One complete page per real question. Google's scaled content policy treats near-duplicate variants as spam, and models understand that "cheapest analytics tool" and "low-cost analytics" are the same question.

Can I pay to be recommended?

Not on the assistants covered here. Ads exist on some of them, labelled as ads. Organic recommendations come from retrieval, and retrieval comes from readable pages and real mentions.

What is the single most useful page to publish first?

An honest comparison of the tools in your category, including yours, with prices you checked and a date. It answers the most common question, it earns links from people who cite it, and it is the page the fan-out search for "best X" returns.