How to get recommended by ChatGPT, Claude, Gemini and Perplexity
When a buyer asks an assistant what to buy, the answer names a few brands and stops. This is what decides which brands those are, what you can change, and how to tell whether it worked.
How an assistant actually picks the brands it names
Three different things happen depending on the question, and they fail for different reasons.
| Path | What the model uses | Why you are missing |
|---|---|---|
| From memory | What was in its training data, months or years ago | You were too small, too new, or barely written about |
| From live search | Pages it retrieves and quotes while answering | Nothing retrievable says what you sell, at what price, for whom |
| From a shopping feed | Structured product data the platform ingests | You publish no feed a machine can read |
Only the first is out of your hands in the short term. The other two are publishing problems, and they are the ones that move within weeks.
The eight things to publish
In the order we do them, because each one makes the next more effective.
Let the assistants in
Check robots.txt for blocks on GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, Google-Extended, PerplexityBot and CCBot. Many stores block them by accident through a security plugin or CDN rule, then wonder why they are invisible. Decide deliberately: blocking training crawlers while allowing search crawlers is a legitimate choice, but blocking the search crawlers costs you the answer.
Say what you sell in plain sentences
Models quote sentences, not layouts. A page that says "premium quality, unmatched service" gives an assistant nothing to repeat. "We roast Ethiopian naturals in 250 g bags, ship within 48 hours of roasting, from £9.50" gives it four quotable facts.
Publish an llms.txt
A short Markdown file at /llms.txt that introduces the site and links the pages that matter. It is a convention, not a standard, and not every engine fetches it. It costs an hour and gives retrieval a clean map of the site. Ours is here.
Give every page a Markdown twin
Serve the same content as clean Markdown at /page.md, linked from the HTML. Retrieval pipelines strip your navigation, scripts and cookie banners anyway; doing it yourself means they quote what you wrote rather than what survived their parser. This page has one.
Mark up the facts with schema
Organization, Product with real price and availability, FAQPage, and Article where it applies. Assistants and shopping surfaces read JSON-LD directly. Prices without a currency, or stock that contradicts the page, are worse than no markup.
Publish a product feed a machine can read
Name, description, price, currency, availability, GTIN or SKU, image, URL. This is what feeds agentic shopping. A catalogue that only exists inside a JavaScript storefront does not exist for an assistant.
Answer the comparison question
Assistants love comparisons because the question is usually a comparison. If you will not write "X versus Y", an affiliate blog will, and the model will quote them describing you. Write it honestly, including who you are not right for.
Earn mentions where models look
This is the uncomfortable one. Most of what an assistant says about you comes from other people's pages: forums, review sites, documentation, news, curated lists. Own-site work makes you quotable and correct; third-party mentions make you recommended. Budget for both.
What does not work
Hidden text aimed at models. Prompt-injection text in white-on-white divs gets stripped, ignored or penalised, and it reads as fraud when someone views the source.
Mass-generated "AI-optimised" pages. Answer engines deduplicate aggressively. Fifty near-identical pages give them one weak source, not fifty.
One-off fixes. Engines re-crawl and change their answers on their own schedule. What you published in March is not what they are quoting in September.
Buying "guaranteed AI ranking". Nobody controls what a model says. Anyone who guarantees placement is either lying or paying for ads and calling it something else.
How to measure it, honestly
The only measurement that means anything is asking the engines the questions your buyers ask, repeatedly, and counting.
- Write 20 to 40 real buying questions in your category, in a buyer's words. Not "best coffee" but "which UK roaster ships beans within days of roasting".
- Ask all four engines — ChatGPT, Claude, Gemini and Perplexity — and record which brands are named, in what order, and what the model cites.
- Count your share of answers, weighted by position. That number is your baseline. In a competitive category, moving from zero to the thirties means going from absent to routinely considered.
- Re-run on a schedule, weekly or monthly, on the same questions. Without the same questions each time, you are measuring the weather.
Our free scan does exactly this for one brand, and shows the fixes it found. Three worked examples with real baselines are here.
Where we stand on rights
Being readable to an assistant does not mean giving up terms. We publish PARP, an open protocol that signs a statement of what agents may do with a page — retrieval, training, commercial use, attribution — bound to the exact bytes served. It is a policy signal rather than enforcement, and it is free to implement. The spec is on GitHub.
Do it yourself, or have it done
Everything above is public knowledge and you can implement all of it by hand. Our free plugins for WordPress and OpenCart publish the machine layer automatically, our software tracks the score, and our team does the work for brands that would rather not. Pricing is on one page.
Questions we get asked
How long until an assistant recommends my brand?
The machine-readability half moves in days: publish the files, the engines fetch them on their next crawl. The recommendation half is slower, because it depends on retrieval and on third-party mentions. In the cases we have measured, meaningful movement in who gets named takes 30 to 90 days, and it is never guaranteed.
Is this SEO?
It overlaps, but the target is different. SEO competes for a position in a list of links. Here there is no list: the assistant names three or four brands and stops. Good technical SEO helps because it makes you retrievable; it does not decide what the model says about you.
Does llms.txt actually work?
It is a convention, not a standard, and adoption by the engines is uneven. We publish it because it is cheap, it makes a site easier to retrieve cleanly, and it costs nothing if an engine ignores it. Treat anyone who calls it a ranking factor with suspicion.
Can I stop AI from using my content and still be recommended?
Partly. You can allow retrieval, which is what puts you in answers, while withholding training. That distinction is exactly what robots.txt cannot express and why we built PARP.
What does it cost?
The scan is free. Self-serve software starts at $49 a month, and managed programmes where our team does the work start at $300 a month. Everything is on one pricing page.