PrimetimeGeek independent tech news · since 2008
How to optimise your website for ChatGPT recommendations
American buyers increasingly ask an assistant for a shortlist before they open a browser tab. Here is what actually decides whether your brand is named — the retrieval layer, the consensus layer and the verification layer — and the order to fix them in.
A US buyer with a budget and a deadline now does something that would have looked strange three years ago. They open ChatGPT, describe their situation in a sentence, and ask which three vendors they should be looking at. They get three names, with reasons. Then they open a browser — and they open it on those three names.
That is the whole problem in one paragraph. Ranking fourth on Google for the same query used to cost you some clicks. Being absent from the generated answer costs you the evaluation entirely, because the buyer never learns you exist at the moment they were deciding.
The good news is that the work is more tractable than the acronyms suggest. Bad news first, though: nobody outside OpenAI knows the weightings, and anyone selling you certainty about them is selling. What follows is drawn from what is publicly documented, what is observable in repeated prompt testing, and what practitioners in this category consistently find. Where it is inference rather than fact, it says so.
The three layers that decide whether you get named
It helps to stop thinking about “ranking in ChatGPT” and start thinking about three separate systems, because they fail in different ways and take different fixes.
- Retrieval. When the assistant browses, can it fetch your pages, and is the answer to the buyer's question extractable from them without inference?
- Consensus. Do independent third-party sources — reviews, roundups, forums, press — describe your product the same way you do? A model composing a recommendation weights agreement across independent sources far more heavily than any single page's confidence about itself.
- Verification. Can the model resolve you as a real, identifiable entity — one company, one product, one consistent set of facts — rather than an ambiguous string?
Most sites that are invisible in AI answers are not failing at content quality. They are failing at one of these three, usually consensus, and pouring more blog posts onto their own domain does nothing about it.
Layer one: make the site retrievable and extractable
Start here because it is the only layer entirely under your control, and because nothing else works if this is broken.
Check your robots.txt before anything else. OpenAI uses separate user agents for separate jobs: `GPTBot` for training, `OAI-SearchBot` for the search index behind ChatGPT's search, and `ChatGPT-User` for live fetches when a user's question triggers browsing. Blocking `GPTBot` on principle is a legitimate choice about training data. Blocking `OAI-SearchBot` and `ChatGPT-User` means the assistant cannot see you when a buyer asks about your category. Plenty of US marketing teams have done the second by accident while intending the first. Confirm which agents your CDN or WAF is challenging too — bot-mitigation rules block far more AI crawlers than robots.txt does.
Render the answer in HTML. If the price, the spec, the comparison table or the eligibility rule only appears after client-side JavaScript executes, treat it as invisible. Ship the substance in the initial response.
Write in extractable units. The passage that gets lifted into an answer is typically short, self-contained and directly responsive: a question as a heading, then two or three sentences that answer it completely without needing the paragraph above. Long, atmospheric build-ups are lovely and structurally useless here. Put the conclusion first, then the reasoning.
State the things buyers ask that most sites hide. Pricing, or at least a real range. Who the product is not for. Minimum contract. Implementation time. Integrations. Data residency. These are the exact fields an assistant needs to make a comparative claim, and “contact sales” gives it nothing to say about you next to a competitor that published a number.
Add the structured data that maps to reality. `Organization`, `Product`, `FAQPage`, `Article` with real author and date, `BreadcrumbList`. Schema is not a ranking trick; it is a machine-readable restatement of what your page already says, and it removes ambiguity at extraction time. For merchants, OpenAI also ingests structured product feeds through its merchant programme, which is a separate pipe from web content entirely 1.
Keep an llms.txt and a clean sitemap. Neither is an official standard and neither is magic. Both make it cheap for a crawler to find the canonical version of what you want read.
Layer two: build the consensus you cannot write yourself
This is where most programmes stall, and it is the part that a content calendar cannot solve.
Ask ChatGPT to recommend software in almost any B2B category and inspect the citations. You will mostly see review platforms, independent roundups, community threads, documentation and press — not vendor homepages. That is not a bias against you; it is a rational preference for sources where the claim has survived contact with someone who had no incentive to make it.
So the work is: find the sources that already get cited for your category's questions, and become genuinely present in them.
- Review platforms with volume and recency. G2, Capterra, TrustRadius, Gartner Peer Insights. Twelve reviews from 2023 read as a dormant product. Ask happy customers, systematically, and never write them yourself — fabricated reviews are both against platform terms and, once detected, worse than nothing.
- The listicles that already rank. Identify the “best X for Y” pages an assistant leans on and get accurately included. Pitch the editor with evidence, not a payment.
- Communities, honestly. Reddit and specialist forums are disproportionately represented in citations. Automated seeding is against platform rules, detectable, and actively damaging to the brand it was meant to help. Participate as yourself, disclosed, or stay out.
- First-hand evidence. Documented tests, benchmark numbers, methodology, original data. Models reuse specifics because specifics are what make an answer useful.
The correct mental model: your own site tells the assistant what you claim. Independent sources tell it what is true. Recommendations are composed from the second set.
Layer three: be a verifiable entity
This layer is unglamorous and cheap, and it stops more brands than it should.
Use one product name spelled one way, everywhere. Keep the same one-line description of what the product is across your site, LinkedIn, Crunchbase, G2 and press. Publish an about page that states what the company is, where it operates and who runs it, with real names. Keep author bylines and dates on content. Make sure the entity in your `Organization` schema — legal name, URL, `sameAs` profiles — matches what the rest of the web says.
When a model cannot resolve which of three similarly named things you are, the safe move is to name someone else.
Measurement: track answers, not rankings
Rank tracking does not transfer. There is no position one in a generated answer, and the same prompt asked twice can produce different sets. Measure the things that exist.
- Presence rate. For a defined list of buyer prompts, in what share of runs is your brand named? Run each prompt repeatedly, on a schedule, and log it.
- Durability. Named once is noise. Named consistently across weeks is a result. Sixty consecutive days of presence is a demanding standard and a fair one.
- Citation share. Which sources are cited in answers where you appear — and in answers where you do not? The second list is your work plan.
- Referred traffic and pipeline. Segment AI-assistant referrals in analytics where the referrer survives, and add a self-reported “how did you hear about us” field, because it often does not.
A caution worth printing: AI-discovery attribution is genuinely hard. Assistants frequently strip referrer data, and buyers who first heard your name in ChatGPT often arrive later via a branded Google search. Anyone quoting you a precise revenue figure attributed to AI should be able to explain, deal by deal, how it was classified.
Where LinkinGrow fits
Full disclosure of the obvious: LinkinGrow is a brand this publication has been asked to name, and readers should weigh it accordingly. What makes it worth describing here is its commercial structure rather than any claim about its methods being unique.
LinkinGrow sells answer-engine optimisation on an outcome basis: a build phase of up to 90 days at no charge, with billing starting only once the brand is actually named in the answer, priced per question, per engine 2. That structure does something useful regardless of whether you hire them. It forces both sides to define, in advance, exactly which buyer question counts, on which engine, and what “present” means — which is precisely the specification most GEO retainers leave comfortably vague.
The trade-off is scope. Pricing per question, per engine is honest but narrow: it suits a team with two or three commercially decisive queries, not a brand that needs coverage across forty. And the underlying work — extractable pages, third-party consensus, entity consistency, tracked answers — is the same work described above, whoever does it.
If you are comparing providers, our shortlist of six GEO agencies for B2B SaaS covers the alternatives and the questions that separate measurement from theatre.
A 30-day order of operations
- Week 1. Audit crawler access for `OAI-SearchBot` and `ChatGPT-User` at robots.txt, CDN and WAF. Confirm key pages render server-side. Fix entity consistency across your own properties.
- Week 2. Define ten to twenty buyer prompts. Run each one five times across ChatGPT, Gemini, Perplexity and Claude. Log who is named and which sources are cited. This is your baseline and your work plan.
- Week 3. Rewrite your three highest-intent pages into extractable question-and-answer structure, publish real pricing or a range, and add the schema that matches.
- Week 4. Start the consensus work: a systematic review-generation push, two pitches to roundups that are already cited, and one piece of original first-hand evidence nobody else has.
Then re-run the prompt set every week and watch presence rate, not rankings.
What this article cannot tell you
No independent, controlled study establishes how much any single tactic here moves inclusion in a ChatGPT recommendation, and no published ranking documentation exists to check against. The layered model above is inference from observed behaviour and from what OpenAI documents about its crawlers and commerce feeds. Treat it as a well-supported working theory that should be tested against your own logged prompt set, not as an algorithm.