A client rang me last spring genuinely rattled. Their organic traffic was flat, which was survivable, but a prospect had told them that when he asked ChatGPT for the best suppliers in their category, three competitors came up and they did not. He wanted to know what to buy to fix it.
There is nothing to buy. There is, however, quite a lot to understand, and almost none of it is what the people selling “AI SEO packages” are telling you.
Retrieval and generation are two different jobs
The single most useful thing you can hold in your head is that an answer engine does two separate things, and they have different rules.
First it retrieves. Given your question, some system goes and finds a set of candidate documents or passages. Depending on the product this is a live web search, a vector search over an index, a cached corpus, or some combination. Perplexity runs live searches. ChatGPT search runs live searches. Google’s AI Overviews sit on top of Google’s own index. Gemini has its own. They are not the same pipeline and they do not always agree.
Then it generates. The model reads the retrieved passages and writes an answer, choosing which of them to lean on and which to cite.
Almost every piece of “AI visibility” advice you have read collapses those two steps into one, which is why so much of it is vague. Being retrievable and being citable are different problems with different fixes. You have to solve both, in that order, because you cannot be cited from a set you were never pulled into.
Getting into the retrieval set
The uncomfortable news for anybody hoping AI search resets the board: classic search fundamentals still do most of the work here. Not all of it, but most.
You have to be crawlable and renderable. If your content only exists after JavaScript runs, some crawlers will see it and some will not. The AI crawlers in particular are less patient than Googlebot. A site that renders its main content server-side is simply in more retrieval sets than one that does not. This is the least glamorous item on the list and the one that most often turns out to be the actual problem.
You have to be in the underlying index. For AI Overviews that means Google’s index and, broadly, Google’s ranking. If you are not in the top ten or twenty for a query, you are unlikely to be in the passage set that gets summarised. AI Overviews are not a way around ranking. They are a layer on top of it.
You should let the crawlers in. Check your robots.txt for GPTBot, PerplexityBot, ClaudeBot, Google-Extended and the rest. Plenty of sites blocked these in a 2023 panic about training data and never revisited the decision. Blocking a training crawler is a defensible choice. Blocking the crawler that fetches pages to answer a live question is a decision to be invisible, and the two often share a name. Know which one you are blocking and why.
Being mentioned elsewhere matters more than it used to. When a model has been trained on a corpus in which your company is described, compared and listed, it has a representation of you that exists independently of your website. That is why brands with strong third-party presence, listicles, comparison pages, forum threads, trade press, get named in answers even when their own site is mediocre. You cannot control that directly. You can influence it by being the sort of company people write about, and by making sure the places that do list your category have you in them correctly.
Getting chosen once you are in the set
This is where things genuinely differ from ten blue links, and where most of the available upside sits.
The model is not reading your page. It is reading a chunk of your page, usually a few hundred to a couple of thousand characters, pulled out of context. Then it decides whether that chunk answers the question well enough to build on and attribute.
Everything below follows from that one fact.
Answer the question in the passage, not across the page
A page that builds a careful argument over eight paragraphs and lands the conclusion at the end is well written and badly chunked. The chunk containing the conclusion has no context. The chunks containing the context have no conclusion. Neither is a good citation.
The fix is not to dumb it down. It is to make each section self-sufficient. State the answer near the top of the section, then support it. Repeat the subject noun instead of leaning on “it” and “this” across paragraph boundaries, because a pronoun three chunks away from its antecedent is a dead reference. If a section would not make sense read aloud on its own, it is not going to get cited on its own either.
Put the specifics in text
Models cite passages that contain something checkable. A number, a date, a version, a named constraint, a price, a threshold. “Improves performance significantly” is not citable. “Cut the largest contentful paint from 4.1 seconds to 1.6 by deferring the chat widget” is, because it says something a summariser can carry over without inventing anything.
This is also why so much marketing copy performs badly in AI answers despite ranking fine. It was written to be persuasive rather than informative, and there is nothing in it to quote.
Structure so the boundaries fall in sensible places
Headings are chunk boundaries in most practical implementations. A descriptive H2 followed by a direct answer is close to the ideal shape. Tables survive chunking well because they carry their own labels. Definition lists work. Walls of text with no internal structure get sliced arbitrarily and the pieces are worth less than the whole.
Give the machine-readable version too
Schema markup is not a ranking factor for AI answers in any way anybody has demonstrated. What it does is remove ambiguity. FAQPage, HowTo, Article with a real datePublished, Organization with a sameAs list connecting your profiles: none of this makes you special, all of it makes you unambiguous. Ambiguity is what gets you left out of a comparison in favour of a competitor whose details were easier to establish.
Say when it was written and when it changed
Recency is weighted heavily, particularly by the live-search products, and particularly for anything that looks like it could have changed. An undated page competing against a dated one loses on a question with any time sensitivity at all. Put a visible published date on the page and a real dateModified in the schema, and only touch the second one when the content actually changed.
What does not work
Stuffing an “AI optimisation” section into the footer. Models are not looking for a magic block. There is no <ai-summary> tag that anybody honours.
Writing for the model instead of the reader. Every piece of this that works also makes the page better for a person: clearer structure, specific claims, stated dates. That is not a coincidence, it is the whole mechanism. Retrieval systems are approximating human judgments of usefulness. Optimising against the approximation and away from the thing it approximates is a losing trade with a short shelf life.
Publishing more. Volume was a reasonable strategy against an index that rewarded coverage. It is a poor strategy against a system that pulls one passage. Ten pages that each half-answer a question are worth less than one that answers it completely, because the ten compete with each other for the same retrieval slot and none of them is the best passage available.
Paying for guaranteed placement. Nobody can sell you a citation. If someone claims to, ask them to show you the mechanism.
Knowing whether it is working
This is the part most people skip, and then a year later they cannot tell you whether any of it helped.
Start by writing down the twenty questions a real buyer would ask on the way to choosing someone like you. Not keywords, questions, in the words a person would use out loud. Then run them, monthly, across ChatGPT, Perplexity, Gemini and Google AI Overviews, and record three things: whether you were mentioned, whether you were cited with a link, and which of your competitors turned up. A spreadsheet is fine to start. There are tools that automate it and some of them are decent, but the discipline of reading the answers yourself for the first few months is worth more than the tooling, because you will notice things a dashboard will not, like the fact that you keep getting described using a positioning you abandoned two years ago.
Watch your server logs too. GPTBot, PerplexityBot, ClaudeBot and OAI-SearchBot all identify themselves. Whether they are crawling you at all, and which pages they take, is the most direct signal available, and it is free.
Then accept that the numbers will be small and lumpy. Referral traffic from these products is a fraction of search traffic and probably always will be, because a good answer often means the person never clicks. The value is not the session. It is being one of the three names that comes up when somebody asks who they should call, which is the same value a good reputation has always had. It is just being brokered by a different intermediary now.
The short version
Be crawlable without JavaScript. Let the right bots in. Rank, because retrieval still mostly runs on the old index. Then write in self-contained sections that answer their own heading, put real specifics in the text, mark up what you can, date everything honestly, and measure by asking the questions your buyers ask.
None of that is a new discipline. It is the old one, done properly, for a reader who now only ever sees one paragraph of your page at a time.