How to show up in AI search.

It retrieves, then it selects. When a buyer asks ChatGPT, Perplexity or Google’s AI Mode a question, the system runs searches of its own, often several at once, pulls a handful of pages into its evidence, and writes an answer from them. Your company can fail at four separate gates:

• Access: can the system reach and read your page at all?
• Retrieval: does your page make the evidence set for this question?
• Evidence use: does the answer keep your claim, and credit it?
• Recommendation: do you actually fit what the buyer asked for?

Each gate fails for different reasons and needs a different fix, which is why a single “AI visibility score” misleads. A mention is not a citation, a citation is not a recommendation, and a recommendation is not revenue. An answer can cite your page while recommending your competitor.

Much of the evidence isn’t yours, either. In Profound’s analysis of 11.84 billion citations, about 57% pointed to company websites, yours and your competitors’, and the rest to editorial coverage and platforms like Reddit and YouTube. So what an assistant says about you is often decided off your website. Whether you call the work GEO or AEO matters less than knowing which of those outcomes you’re losing. For the numbers behind all this, each checked against its source, see our AI search statistics.

Usually for a mechanical reason before a strategic one. Work down the causes in order, because each one hides the ones after it:

• It can’t fetch you. Robots rules, bot filtering at your CDN, or copy that only appears after JavaScript runs.
• It can’t find you. ChatGPT’s search tracks Bing closely: in one 500-citation sample, 87% of SearchGPT citations matched Bing’s top results. If you’re missing from the index an engine leans on, you start from behind.
• It can’t tell who you are. A vague homepage, a common name or conflicting facts across the web make you hard to resolve as one company.
• It has better evidence for someone else. Your competitor’s pages prove what yours only assert, or the sources it trusts mention them and not you.
• They really are the better fit. Then no content will fix it, and none should.

Two instincts make this worse. Google rankings don’t transfer: only 38% of AI Overview citations came from pages in the organic top 10 in March 2026. And when an answer is flat wrong about you, there’s no form to file with the model. You correct the sources it reads.

Ask it, the way your buyers do, and keep the receipts. Write down the questions a real buyer asks: what you do, who you’re for, how you compare with a named competitor, whether you fit a specific situation. Run each one in a fresh chat with search on, and save the date, the full answer and every cited link.

One run proves nothing. Answers change from run to run, and recommendation lists rarely repeat exactly, so ask every question in two or three phrasings, on at least two engines, every week. Report counts, like “named in 6 of 10 answers”, not a verdict from one screenshot.

Then freeze the set. About thirty questions with fixed wording, rerun on a schedule, beat hundreds of loose ones, because a question you’ve reworded can’t tell you whether the answer changed. Carry in the constraints buyers really mention, like budget, company size or an integration. In real conversations, those often sit in earlier messages rather than the final question.

Fix eligibility before you touch a sentence. For Google’s AI features, a page has to be indexed and eligible for a snippet, and Google says nothing more is required: no special markup, no new files. The rest of the engines are just as unforgiving about access. Check these on the pages your buyers need most:

• Crawler access. OpenAI alone runs four crawlers with separate jobs, and allowing one doesn’t allow the others. Check your CDN too: some block AI crawlers by default.
• Real HTML. The major AI crawlers in Vercel and MERJ’s study fetched JavaScript files without running them; Google does render. If your copy only appears after scripts run, those crawlers see an empty page.
• Clean responses. A 200 status, no accidental noindex, one canonical page per buyer job, and URLs that don’t break or chain redirects.
• The pages agents need. Plain comparisons, pricing logic, integrations and outcome proof answer the questions a homepage can’t.

Schema and llms.txt are hygiene, not levers. No published study ties either one to citations, so give them an hour, not a quarter.

Evidence, not phrasing. The study that named GEO found that adding statistics, quotations and cited sources raised a page’s visibility inside generated answers by up to 40% on its own benchmark, while keyword stuffing produced no gain. A larger retest, C-SEO Bench, found most rewrite methods largely ineffective and often harmful to citation rank, and gains shrank as more sites adopted them. Tricks get copied. Proof is harder to copy.

So publish what only you can: an original number, a named method, a result with its denominator. Write paragraphs that stand on their own, with the answer first, then the support, then the conditions that keep the claim true. Models use what sits at the start of their context more reliably than what’s buried in the middle (Liu et al.), so lead with the point.

Then earn it where the answers look. Much of what engines cite sits off your domain, so real reviews, accurate listings and data others want to cite do more than another blog post. Never manufacture any of it. And drop the scarcity copy: in one study, scarcity and exclusivity framing lowered how often a model recommended a product.

Measure each outcome on its own, against your own noise. Mentions, citations, comparisons, recommendations, traffic and revenue move for different reasons and call for different fixes, so blend them into one score and you lose the reason. The IAB’s guidance makes the same split, into presence, prominence, portrayal and persuasion, and calls any program under 50 queries exploratory.

Before you celebrate or panic, check the spread. Run-to-run noise, churn in what the engine retrieves and a real model update look identical from a single check. Compare the same frozen questions, run the same number of times, over several weeks.

Traffic is the hardest part. Many AI visits arrive without a referrer and land in Direct, and Google’s AI Overviews and AI Mode report inside Organic Search, so any AI channel you build is a floor, not a census. Treat revenue as a chain you join one step at a time, and label each claim as observed, attributed or causal.

Spend in the order the evidence says, cheapest check first. Ninety days looks like this:

• Days 1–30: write 15 to 25 buyer questions, two or three ways each, and run them weekly on two engines. Fetch your ten most important pages with JavaScript off, check your crawler rules, and fix what that turns up.
• Days 31–60: close the gaps the baseline names, mechanical ones first. Then build the comparison and proof pages buyers and agents keep asking for.
• Days 61–90: rerun the identical questions. One mention in ten becoming two is noise. One becoming six, and holding across reruns, is a signal.

On a small budget, pick one narrow territory where nobody owns the answer yet, instead of chasing the broad category term. Ninety days buys a baseline and a readable site. It doesn’t buy a guaranteed citation, and nobody can honestly promise you one.

If you buy software, buy it for the job: tracking answers, diagnosing sources, producing evidence or checking retrievability. Ask every vendor how many runs sit behind each number it shows you.