In December 2024, OpenAI reported that ChatGPT had passed 300 million weekly active users. A meaningful share of those people now ask a model for a recommendation before they ask a search engine, and many never reach a traditional results page at all. They type “what is the best CRM for a 10-person agency” or “compare these three accounting tools for a freelancer,” read the answer, and act on it. The recommendation happened. Your website was never involved.
This is the shift that should reorganize how you think about visibility. For 20 years, the goal was to rank a page so a buyer would click it. Now a buyer can complete the research stage of a purchase without clicking anything, because the model did the comparison and named the winners. If your product was not in that answer, you did not lose a ranking. You lost the consideration set, which is worse, because the buyer never knew you were an option. The work now is to influence AI recommendations directly, by changing the signals the model reads when it builds the answer. There are five of them.
Why AI recommendations happen before your website does
A language model does not browse the way a person does. When someone asks for a product recommendation, the model assembles an answer from what it already knows plus, in the case of engines with live retrieval, what it can pull from the web in that moment. It is not loading your homepage and forming an impression. It is checking whether a coherent, corroborated picture of your product exists across the sources it trusts, and whether that picture matches the question being asked.
That changes where the buying decision gets shaped. The old model put your website at the front of the journey: the buyer searched, clicked, and you made your case on your own page, in your own words. The new model puts the AI engine at the front. By the time a buyer reaches your site, they have often already been told what to consider, and your product is either on that list or it is not. To influence AI recommendations, you have to move your effort upstream, to the sources and signals the model reads before it ever mentions you. The website still matters, but it has become the place the model verifies facts, not the place the buyer forms a first impression.
The five-signal model
Here is the framework. When an AI engine decides which products to recommend, it is weighing five signals. I call this the five-signal model, and the value of naming it is that it turns a vague goal, “show up in AI answers,” into five concrete things you can audit and fix.
Signal one is entity clarity: whether the model has a single, consistent understanding of what your product is, who makes it, and what category it belongs to. Signal two is third-party corroboration: whether independent sources, reviews, comparisons, directories, and articles say the same things about you that you say about yourself. Signal three is structured specifics: whether concrete, extractable attributes exist, price, use case, integrations, capacity, the facts a model needs to match you to a specific question. Signal four is recency: whether the signals about you are current, because a model weighing two options will lean toward the one with fresher corroboration. Signal five is consensus language: whether you are described using the words the category actually uses, so the model connects you to the query.
A product strong on all five gets recommended often and accurately. A product weak on two or three gets skipped, or worse, gets described wrong. The rest of this piece works through the signals in the order they tend to fail.
Signal 1 and 2: who you are, and who agrees
Entity clarity fails more often than anything else, and it fails quietly. The model encounters your product described as a “marketing platform” on one page, a “content tool” on another, and a “workflow app” on a third. Each description is defensible. Together they prevent the model from forming a confident answer to the question “what is this,” and a model that cannot answer that question will not stake a recommendation on you.
Fixing entity clarity means picking one precise description of what you are and what category you compete in, then making every surface you control say it the same way: your homepage, your About page, your social profiles, your directory listings, your structured data. Consistency is the signal. The model is not impressed by clever positioning; it is reassured by repetition.
One overlooked place entity clarity breaks is your own structured data. If your site’s metadata, your schema markup, and your social profile descriptions each label you a little differently, you are feeding the model contradictions through the channels it reads most literally. A model can forgive a marketing page that gets poetic about what you do. It is far less forgiving of structured fields that plainly disagree with each other. Audit the machine-readable description of your product everywhere it exists, and make every instance say the same plain thing in the same plain words.
Signal two, third-party corroboration, is where most of the real work lives, because it is the signal you cannot fake. The model trusts a claim more when sources you do not control confirm it. If your site says you are the best option for small agencies, that is marketing. If three independent review sites, a comparison article, and a handful of forum threads also say you work well for small agencies, that is corroboration, and corroboration is what the model recommends on. You build it the slow way: earning genuine reviews, getting included in legitimate comparison content, being mentioned accurately in articles and directories. Every corroborating source is a vote, and to influence AI recommendations you need enough votes that the model sees a pattern rather than a single self-interested page.
Signal 3, 4, and 5: specifics, recency, and consensus language
Structured specifics is the signal that decides whether you get matched to a precise question. A buyer rarely asks “what is a good CRM.” They ask “what is a good CRM under 50 dollars a month that integrates with QuickBooks and works for a team of five.” A model can only place you in that answer if the concrete attributes exist in extractable form: the price, the integration, the team-size fit, stated plainly somewhere the model can read them. Vague benefit language, “affordable,” “integrates with the tools you love,” gives the model nothing to match. Specific, factual statements do. Audit your own pages and ask whether a model could answer ten common buyer questions about your product using only the facts you have published. If it cannot, you have found your gap.
There is a fast way to run that audit. Write down the ten questions a real buyer asks on the way to choosing a product like yours, the practical ones about price, fit, integrations, capacity, and limits. Then try to answer each one using only text that sits on a page a model could read. The questions you cannot answer that way are your content gaps, and they come pre-ranked, because some of those questions get asked far more than others. Close the most common ones first, in plain factual language, and you have done more for your AI visibility than another round of homepage copy ever would.
Recency is the simplest signal to act on and the easiest to ignore. Models lean toward the option with current corroboration, partly because freshness is a proxy for “still exists and still works.” A product whose most recent independent mention is two years old looks dormant next to a competitor reviewed last month. You do not need a constant firehose, but you do need a steady pulse of new, accurate, third-party signal so the model never has to choose between you and a fresher rival.
Consensus language is the signal people miss because it feels like a branding decision. If your category calls the thing “appointment scheduling software” and you have spent three years calling your product a “client booking experience,” the model may not connect you to the query at all, because the buyer asked using the category’s words and you answered in yours. You can keep a distinctive brand voice. You also need to plainly state, somewhere the model reads, the standard category terms a buyer would use. Speak the category’s language for the match, then use your own voice for the persuasion.
Consensus language also decides which comparison set you land in. AI engines sort products into categories before they rank within them, and the sorting runs on language. If the model cannot tell from your words that you belong in a category, it will not consider you when a buyer asks for the best option in that category, no matter how good the product is. The category term is the ticket into the room. Your distinctive voice is what you do once you are inside it, and a clever name for your category is worth nothing if it keeps you standing outside the door.
Where to start this week
Do not try to fix all five signals at once. Start with an honest audit. Open the AI engine your buyers use most and ask it the three or four questions a real prospect would ask on the way to choosing a product like yours. Read what it says. Note whether you appear, whether the description is accurate, and which competitors it names instead.
That exercise tells you which signal is failing. If you do not appear at all, the problem is usually entity clarity or consensus language: the model does not have a confident picture of what you are or does not connect you to the category term. If you appear but the description is wrong, the problem is corroboration or specifics: the model is working from thin or stale sources. If you appear but always behind the same two competitors, the problem is usually corroboration volume and recency: they have more independent votes, and fresher ones.
Whatever the audit shows, resist the urge to fix all five signals at a shallow level. A model rewards genuine depth on one signal over a thin pass at all five. A product with abundant, real third-party corroboration and a slightly messy category description still gets recommended. A product with tidy metadata and almost no independent corroboration does not. Find the one signal failing hardest, make it genuinely strong, and only then move to the next. The five-signal model is a diagnosis tool, and a diagnosis is useful only if it sends you to treat the actual problem.
Pick the one failing signal and spend the next quarter on it. The buyer who once started on a search results page now starts inside an answer, and the answer is being written whether you influence it or not. Run the audit today, find your weakest signal, and start there.