AI Search Strategy: the four pillars that decide who gets mentioned

For twenty years the job was to rank on a list. A buyer typed a query, got ten links, and chose between them. Now the buyer asks an assistant a question and gets one answer. You are either inside that answer or you are not in the conversation, and there is no second page to be on.

This is not a small channel arriving slowly. At Google I/O in May 2026, Sundar Pichai put AI Overviews at over 2.5 billion monthly users and AI Mode at over 1 billion within its first year. OpenAI reported 900 million weekly active ChatGPT users in February 2026. Your buyers are already there, and the assistant is already naming somebody.

The number that should actually move a founder is not the traffic figure. It is what the answer does to a purchase decision. G2 surveyed 1,076 B2B software buyers in March 2026 and found that 69% chose a different vendor than the one they had planned on, based on what an assistant told them, and 33% bought from a vendor they had never previously heard of. Being in the answer is not a visibility metric. It decides who gets bought.

What follows is four pillars. They are sequential for a reason: each one is wasted effort until the one above it is true. Under each is ahow to get started block you can act on this week.

Pillar 1 Make yourself readable to a machine

Most businesses respond to weak AI visibility by publishing more. Nothing changes, and the conclusion drawn is that the channel does not work. The usual cause is upstream of content entirely: the retrieval layer cannot read the site.

Three gates decide whether citation is even possible.

Gate one is crawler access. Your robots.txt names which bots may read you. GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and their equivalents each need explicit permission. Plenty of sites block them by inheritance rather than decision, because a developer copied a restrictive robots file or a security plugin added one. If you are blocked here, you are invisible by configuration and no amount of content changes it.

Gate two is schema markup. Schema tells a machine what a page is rather than making it infer. Article, author, organization, FAQ, product. A model that already knows a large brand can infer its way to a reasonable answer. A model meeting a mid-market business for the first time infers badly, and inference is where wrong answers about your business come from.

Gate three is a parseable page. Answer first, one claim per sentence, real headings rather than styled div soup. If your strongest fact lives inside an image, a video, or content injected by script after load, it does not exist to a retrieval system. This is the gate that beautiful marketing sites fail most often.

There is a fourth item worth adding now rather than later, because it is cheap and the surface is growing: an llms.txt file and clean markdown versions of your key pages. Agents that act on behalf of a buyer are a different consumer from a model that indexed you months ago. They arrive, read, and decide in one pass. Giving them a plain, structured statement of what you do and what you sell costs an afternoon.

The order here is not arbitrary. Fixing what an assistant already says about you comes before trying to increase how often it says anything. A wrong answer is worse than silence, because it is confidently delivered and it travels. If ChatGPT currently describes a product you retired, or attributes a competitor's pricing to you, that is the first thing to correct, and correcting it usually means publishing an unambiguous, machine-readable page that states the correct fact plainly.

Pillar 2 Map the questions, not the keywords

A buyer prompt does not produce a single lookup. The assistant decomposes it into a set of background retrieval queries, gathers what comes back, and assembles an answer from the results. Those background queries are called fan-out queries, and they are the surface that actually decides citation.

This is why a page can rank well and never get cited. It was written for the phrase a human types. The machine went looking for something adjacent and more specific, found a better match elsewhere, and named that source instead. Keyword tools describe the visible prompt. They do not describe the fan-out.

The practical consequence is that language placement matters more than keyword density ever did. The question a buyer would actually ask belongs in the URL, the title, the H1, and the H2s, in their words rather than your industry's. A page called "What is GEO" is written for practitioners. The same page called "How to get your business recommended by AI search" is written for the person with the problem, and it is the second one the machine matches against.

Here is what the decomposition actually looks like. A buyer asks an assistant "what is the best AI search tool for a B2B software company." That single prompt fans out into questions the buyer never typed: what counts as an AI search tool, how they differ from SEO tools, which ones support B2B specifically, what they cost, what company size each suits, which integrate with an existing stack, and what the alternatives to buying one are. The answer is assembled from whatever best matches each of those, and the brand that gets named is usually the one with a page aimed squarely at one of them rather than at the headline prompt.

That has a useful implication for a smaller business. You do not have to win the headline question to get into the answer. Winning one of the specific questions underneath it puts your name in the same response, and those questions are far less contested. This is the entire reason a challenger can appear alongside an incumbent inside a month.

In the platform

Pillar 3 Publish at a cadence that outruns decay

Content does not age gracefully in the answer layer. The retrieval layer has a recency preference, and a library that is not maintained slides out of the answers quietly while the traffic report still looks acceptable.

That single finding reframes the whole budget question. If freshness is the strongest lever, then a fixed library of any size is a depreciating asset, and the spend that matters is the spend on maintenance rather than on the next new article. Most content operations are built the other way around.

It also explains why this channel is winnable by smaller businesses. An incumbent with a decade of domain authority and a stale library is beatable by a challenger publishing and refreshing continuously, because the retrieval layer is matching a question to the best current answer and not consulting a seniority list. Freshness is the lever an incumbent is least likely to pull, because pulling it requires operational cadence rather than budget.

The arithmetic is where doing it yourself breaks. Covering a mapped question space and keeping it current is a full-time function. In a business with zero to three marketers there is nobody to assign it to, which is the actual reason most AI Search strategies stall after the first month of enthusiasm.

The economics are worth stating plainly, because the comparison is usually made against the wrong thing. A human content agency runs somewhere around $10,000 a month for seven to ten articles, with no refresh loop attached. A managed content engine runs around $5,000 a month and produces both new articles and continuous updates. The second number buys worse prose. It also buys the three things this channel rewards, which are volume, structure, and freshness, and prose quality is not on that list. That is an uncomfortable sentence for anyone who has spent a career on craft, and it is still what the citation data shows.

In the platform

Pillar 4 Measure citations, not clicks

The click used to be both the goal and the measurement. It is no longer either, and continuing to grade this channel on clicks produces a confident, wrong conclusion that it does not work.

Four measurements replace the one you lost. Citation shareis the headline: take your ten buyer questions, run them across ChatGPT, Gemini, Perplexity, and Google AI Overviews on a schedule, and count how often you are named. Branded search is the downstream signal, because a buyer who reads your name in an answer types it into Google next.

AI referrer sessions are the third: segment chatgpt.com and its equivalents separately in analytics, because that traffic converts like a referral rather than like search. Functionally it is a referral. Something recommended you.

The fourth is impression and decay curves in Search Console, which is your early warning that a page is sliding while the revenue impact is still small enough to fix cheaply.

Attach one honest caveat to all of it. A large share of AI-driven demand never carries a trace back to the answer that caused it.

In the platform

A baseline is less work than it sounds. Ten questions, four surfaces, forty prompts. Run them, record for each one whether you were mentioned, cited with a link, or absent, and note who was named instead. That is an afternoon, and it converts a vague worry about AI visibility into a number you can move. Repeat it monthly on the same questions, because the same prompt does not return the same answer twice and a single reading is noise rather than signal.

What to avoid

Three approaches actively cost you ground, and all three are common enough to be worth naming.

Self-serving comparison pages

The "best tools in our category" page where you happen to be first is a genuine own goal here. You have published a structured, machine-readable list of your competitors on your own domain, and the retrieval layer will cheerfully use it to answer questions about them. Write comparisons only where you would be comfortable with the assistant quoting the whole list.

Bulk unreviewed AI content

Volume alone is the failure mode, not automation. Pages generated without question mapping, structure, or a maintenance loop produce a short impressions spike and then a long slide, and they carry real risk under Google's spam policies on scaled content abuse. Automation with a system behind it is a different thing from automation instead of a system.

Blocking AI crawlers by accident

Blocking AI crawlers is a legitimate decision for some publishers. Doing it without knowing is not a decision, it is an outcome. Read your robots.txt, and if you are blocking, be able to say why.

If you only do one thing this month

Run the baseline. Not because measurement is more important than the work, but because almost every business I talk to is wrong about where they currently stand, and the direction of the error is not consistent. Some are already being mentioned and have no idea, which means the job is protecting a position rather than building one. Others are certain they show up because they rank first in Google, and the assistant has never named them once.

Those two businesses need opposite strategies, and no amount of general advice tells you which one you are. Forty prompts does.

Common questions

Is AI Search optimization just SEO with a new name?

No. SEO optimizes for a ranked list a human reads. AI Search optimizes for citation inside a machine-generated answer. The retrieval mechanics differ, the success metric differs, and the authority model differs. SEO earns authority through backlinks. AI Search earns it through topical coverage and freshness.

Will Google penalize AI-generated content?

Google penalizes low-quality content, which it always has. Relevant, structured, fresh, specific content performs regardless of how it was produced. The production method is not the variable being judged. Quality, structure, and freshness are.

How long before a business appears in AI answers?

Coverage and impressions in weeks. First mentions around thirty days from a serious start. Citations typically inside one to three months. Compounding after month three, once topical coverage accumulates and the refresh loop is running.

My competitors already get cited. Is it too late?

No. Relevance and freshness beat tenure in this channel. A challenger targeting specific fan-out queries, comparisons, and situations can outrun an incumbent whose library has gone stale, because the retrieval layer favours recently updated pages.

What has to be in place before any of this works?

Three technical gates. AI crawlers unblocked in robots.txt, schema markup on your pages, and a page structure a machine can parse. If the retrieval layer cannot read your site, nothing downstream matters.

How do you measure AI Search when there is no click?

Track four things together: citation share across ChatGPT, Gemini, Perplexity, and Google AI Overviews; branded search volume; AI referrer sessions segmented in analytics; and impression and decay curves in Search Console. Each one alone is misleading. Together they show the channel.