How to optimise for AI search in 2026: the complete guide
The short answer
Most complete guides to this subject are a flat list of fifteen tactics with no evidence attached and no order.
This one is ordered, and three of the tactics most often recommended have measured evidence against them.
- Adding schema markup does not increase AI citations. A controlled study of 1,885 pages found minus 4.6 percent on AI Overviews and no significant effect elsewhere.
- Publishing llms.txt does almost nothing. Across 137,210 domains, 97 percent of these files received zero requests in a month.
- Publishing more does not reliably help. Updating an existing page is the better bet. Among pages LLMs cite, 75 percent had been updated in the last year, while only 42 percent were recently published.
Meanwhile the lever with the strongest evidence behind it is the one these guides discuss least: what other people publish about you.
Key takeaways
- Retrieval comes first, and it is mostly classic SEO. 88.46 percent of ChatGPT citations came from pages retrieved through the search channel.
- Retrieval is roughly half the job. Of the pages ChatGPT retrieves, it cites 49.98 percent.
- Your title does more than your body copy. Cited pages scored 0.602 on similarity to the prompt, against 0.484 for non-cited.
- Earned third-party coverage carries the most weight. 84 percent of AI citations come from earned media, and 0.3 percent from paid.
- Citation is worth money, and does not restore the old number. With an AI Overview present, cited pages earn 2.07 percent click-through against 0.94 percent uncited. Both sit below the 3.35 percent of a page with no Overview.
Where this comes from. We run organic for more than 250 enterprises at Pepper across 8 years, and track over 10 million prompts on every major engine. Everything below traces to a named study with a published method. We also name what we excluded, and why. That discipline shapes the view here, along with the client reviews we sit in and the conversations we have at the events we run.
What is AI search optimisation?
AI search optimisation is the work of getting an engine to retrieve your page, quote it, and describe your brand correctly. All three have to happen when someone asks a question you should own.
It breaks into three outcomes that move independently. Conflating them is the most common planning error we see.
- Being retrieved. The engine pulls your page into the candidate set. This is mostly classic SEO, and it gates everything else. What a results page now contains explains why that matters more than it sounds.
- Being cited. The engine quotes your page in the answer. This is writing craft, judged at passage level rather than page level, and E-E-A-T covers how engines judge whether to trust it.
- Being named and described correctly. The engine mentions your brand and puts you in the right category. This depends mostly on what other people publish about you.
Each needs different work and a different budget. Brand visibility covers the third in depth.
See where you show up. Pepper’s GEO platform tracks Brand Visibility, Domain Prompt Presence and Share of Voice across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews. Your team can log in, connect Search Console and GA4, manage the prompt set and run your own agents in the Agent Atlas. A growth team works the same account alongside you. Book a growth audit or see where you show up.
The order that matters
Each layer gates the one above it. Working out of order is the most expensive mistake in this field.
Most programmes we inherit have spent heavily on the top layer and skipped the bottom two entirely. That produces beautifully written pages nothing retrieves.
1. Clear the technical floor
Free, an afternoon, and the check most often wrong.
Read your robots.txt against the real crawler list rather than a generic one. Two corrections matter more than the rest, and both come from vendor documentation.
- Blocking GPTBot does not hide you from ChatGPT. OpenAI describes GPTBot as crawling content that may train foundation models. The crawler controlling ChatGPT search visibility is OAI-SearchBot.
- Blocking Google-Extended does not remove you from AI Overviews. Google states it does not impact inclusion in Search or act as a ranking signal.
Then fetch your key pages as raw HTML. If the answer only appears after JavaScript runs, some retrievers never see it. The full method is in the free AI visibility audit, and the schema half is in schema markup and AI search.
2. Earn retrieval, which is mostly classic SEO
This is the step the category’s marketing tends to skip.
Across 1.4 million ChatGPT prompts, citation rate by retrieval channel ran at 88.46 percent for search, 12.01 percent for news and 1.93 percent for Reddit. YouTube reached 0.51 percent and academia 0.40 percent. To be cited, you generally need to sit in the search selection pool first.
So ranking still matters, and more than most AI search advice admits. We set out that argument in full in AEO vs SEO.
3. Make the page quotable
Being retrieved is roughly half the job. ChatGPT retrieves about 33 pages per prompt and cites 49.98 percent of them.
Three measured findings decide which half you land in.
- Title similarity to the question. Cited URLs averaged 0.602, non-cited 0.484. So title your page as the question a buyer types, rather than as your positioning deck phrases it.
- Readable URLs. Natural-language slugs were cited 89.78 percent of the time against 81.11 percent for opaque ones. Free on new pages, rarely worth retrofitting.
- Answer-first passages. Engines score passages against sub-questions. Every section must survive being lifted out and read alone.
The detail is in how to write content that AI engines actually cite.
4. Earn coverage you do not control
The highest-evidence lever, and the least discussed.
84 percent of AI citations come from earned media, against 0.3 percent from paid and advertorial. A separate analysis of 410,000 brand mentions found repeatedly visible brands drew 70 to 80 percent of their mention footprint from third parties. Rarely visible brands were owned-heavy, at 30 to 40 percent.
Two cautions come with it. Roughly half to 60 percent of those mentions carry no hyperlink, so backlink tools cannot see them. The source mix also moves fast, which is why earned authority argues for spread across several source types rather than depth on one.
What the evidence says not to do
This is the section that separates this guide from the rest of the SERP.
Three tactics appear in almost every complete guide to AI search. All three have been tested. None earns its place on the evidence.
Read the limits on each before treating this as settled. The schema study only covered pages already carrying 100 or more AI Overview citations. So it tests whether schema helps already-cited pages, rather than unknown ones. The llms.txt sample skews technical and SEO-aware. And no study can prove a negative. What we can say is that the measured effects are absent, small or negative, while the tactics get recommended as though they were large and positive.
Our methodology: how we ranked the levers
| Lever | Weight | Evidence behind the weighting |
|---|---|---|
| Technical floor and crawler access | 25% | A gate. Vendor documentation shows the common blocks are misunderstood |
| Retrieval through classic search | 25% | 88.46% of ChatGPT citations came via the search channel |
| Page-level citability | 25% | Only 49.98% of retrieved pages get cited, and title similarity separates them measurably |
| Earned third-party coverage | 25% | 84% of citations come from earned media, and mention composition separates visible brands from invisible ones |
What we could not verify and left out. A long list, and naming it is part of the method: the 68 percent zero-click figure, 13x ROI claims, GEO case studies claiming 4,900 percent revenue growth, SERP pixel-share percentages, agent accuracy rates, and the claim that 44.2 percent of citations come from the first 30 percent of a page. None traces to a published method we could open.
AI search optimization levers at a glance
| Lever | Evidence strength | Typical 2026 cost | Time to signal | Where it falls short |
|---|---|---|---|---|
| Crawler access, checked by name | Vendor documentation, unambiguous | Free, an afternoon | Immediate | A gate, not a lever. Wins nothing alone |
| Rendering without JavaScript | Direct, mechanical | Developer time | Weeks | Invisible in most dashboards |
| Classic ranking | 88.46% of citations came via search | Your existing SEO budget | Months | Slowest, and unavoidable |
| Title as the question | 0.602 against 0.484 similarity | Free, one line per page | Weeks | Fails if the page does not then answer it |
| Answer-first structure | Engines score passages, not pages | A day of editing per page | Weeks | Retrofitting an archive is slow |
| Updating existing pages | 75% of cited pages updated within a year | Cheaper than new pages | 1 to 3 months | Reports badly, looks like nothing shipped |
| Earned third-party coverage | 84% of citations from earned media | The largest line item, and slowest | 3 to 6 months | You do not control the outcome |
| Schema markup | Minus 4.6% on AI Overviews in a controlled test | Low | None measured | Still worth having for rich results |
| llms.txt | 97% of files got zero requests | An hour | None measured | Real use case in developer docs only |
| Paid placement | 0.3% of AI citations | Media spend | None for citation | Buys attention, not citation |
How to choose where to start
The binding constraint differs by company, so the honest answer is that you have to measure rather than guess.
The scorecard
Score your programme out of 100.
| Area | Weight | How to score it honestly |
|---|---|---|
| Crawlers verified individually by name | 25% | Someone checked OAI-SearchBot, PerplexityBot and Googlebot separately, not as one line |
| Pages retrieved through classic search | 25% | You rank for the queries you care about, which gates the citation half |
| Titles and answers shaped as questions | 20% | Read a title aloud. Would a buyer type something close to it? |
| Third-party mention composition | 20% | Credible independent sources describe you, and you can name three |
| Named and cited tracked separately | 10% | Two numbers per engine, not one blended score |
The live test, over 90 days
Fix 30 prompts in your buyers’ words and hold them steady. Real questions, such as “how to optimise for AI search”, “best data observability platform for a small team”, “how much does sales enablement software cost” and “which GEO agency works with fintech”.
Run them across Google, ChatGPT and Perplexity. Record four columns. Whether an AI answer appeared, whether you were named, whether a page of yours was cited, and which sources appeared instead.
Rerun the identical set monthly for 90 days. Three readings is the shortest honest window. One is a snapshot, two could be noise, three shows direction. Keep the set fixed, because changing it resets the comparison.
That fourth column is the plan. It is an earned-media target list built from what engines already trust in your category, and it costs an afternoon. The free method is in how to measure AI search visibility without expensive tools.
Weak approach versus strong approach
The weaker approach: add schema everywhere, publish llms.txt, raise output to four posts a week, then buy a dashboard to watch it. Three of those four have measured evidence against them, and the fourth only reports.
The stronger approach: check three crawlers by name and fix rendering. Retitle your twenty most valuable pages as the questions they answer. Then work the source list your own prompts return.
The weak version is a list of tasks that look like work. The strong version is short and unglamorous, and every item traces to a published study.
Red flags
- A guide with no studies attached. This field is full of confident rules nobody has tested. Ask what evidence sits behind each one.
- Schema or llms.txt as a headline lever. Both have been measured, and neither earns a headline.
- “Ranking no longer matters.” Search was the retrieval channel behind 88.46 percent of citations.
- A plan entirely on your own domain. That is the profile of the rarely-visible group in the mention data.
- One blended visibility score. It cannot separate an awareness problem from a citability one.
- Guaranteed citations or rankings. Nobody controls generated output, and Google warns against providers guaranteeing rankings.
Five questions worth asking
- “Which crawlers did you check, by name?” If the answer is “AI crawlers”, nothing useful was checked.
- “What study is this recommendation based on?” Category consensus is a hypothesis, not evidence.
- “Are we tracking named and cited separately?” One number loses the diagnosis.
- “Which sources appear instead of us?” That is the target list, and it is free to build.
- “What would you tell us to stop doing?” A good partner reallocates rather than adds.
Reduced to one principle: clear the floor, earn retrieval, write something quotable, then get other people to describe you accurately.
One closing note that costs us something. Most of the highest-evidence work here is free or cheap. Three crawler checks, a raw HTML fetch, twenty retitled pages, thirty prompts run by hand. The expensive parts, dashboards and volume content, are the easiest to sell and the weakest on evidence, ours included. Do the free half first.
What does AI search optimisation cost?
| What you are buying | Typical 2026 cost | What it covers |
|---|---|---|
| The technical audit | Free, an afternoon | Crawler access by name, indexing, rendering |
| Search Console and Bing Webmaster Tools | Free | Your own first-party data on two surfaces |
| Manual visibility tracking | Free, about two hours a month | 30 prompts, three engines, four columns |
| Retitling and answer-first edits | Roughly a day per page | The citability half |
| Updating existing pages | Cheaper than new production | 75% of cited pages were updated within a year |
| Earned third-party work | The largest line item, and slowest | Where 84% of citations originate |
| Multi-engine tracking | Roughly $99 to $400 a month | Automated, per engine, with history |
| Platform plus a growth team | Custom | Pepper: tracking, plus the people and agents doing the work |
The first three rows are free and cover most of the evidence-backed work.
What nobody should promise you
A guaranteed citation or ranking. Nobody controls generated output, and Google warns against providers guaranteeing rankings.
Citation gains from schema markup. A controlled study measured minus 4.6 percent on AI Overviews and nothing significant elsewhere.
Visibility gains from llms.txt. Across 137,210 domains, 97 percent of these files were never requested.
That being cited restores your old click-through. A cited page with an AI Overview present still trails a page with no Overview by roughly 38 percent.
Historical data from before you started measuring. There is no backfill in this category, which is the honest argument for starting now rather than buying more later.
When not to start here
The answer that costs us the sale. If your pages are not indexed, you do not need an AI search programme yet and we would tell you not to buy one.
Fix indexing first. It is free, it takes an afternoon in Search Console, and it gates every layer above it. An AI search programme on an unindexed site buys a detailed account of why nothing works. You could have had that for nothing.
Frequently asked questions
How do I optimise for AI search?
In order: clear the technical floor, earn retrieval through classic search, make individual passages quotable, then earn third-party coverage. Each layer gates the one above it. Working out of order is the most expensive mistake in this field.
Does schema markup help AI search?
Not for citations, on the best evidence available. A controlled study of 1,885 pages found minus 4.6 percent on AI Overviews, and no significant effect on AI Mode or ChatGPT. It remains worth having for rich results in classic search.
Should I publish an llms.txt file?
For AI search visibility, no. Across 137,210 domains, 97 percent of these files received zero requests in a month. The genuine use case is developer documentation, which is a different job.
Does classic SEO still matter for AI search?
Yes, more than most advice admits. Across 1.4 million ChatGPT prompts, 88.46 percent of citations came from pages retrieved through the search channel. Ranking is the gate on the citation half.
Is it better to publish new content or update existing pages?
Update, usually. Among pages LLMs cite, 75 percent had been updated within the last year, while only 42 percent had been published recently. Freshness is largely manufactured by updating rather than publishing.
How long does AI search optimisation take?
Plan on three monthly readings of a fixed prompt set before judging direction, so roughly 90 days for a signal. The earned third-party half runs slower, because it depends on other people publishing.
What actually drives AI citations most?
Earned third-party coverage carries the strongest evidence, at 84 percent of citations against 0.3 percent from paid. Page-level craft then decides whether a retrieved page gets quoted. Retrieval itself depends on classic ranking.
How do I measure AI search visibility?
Run thirty fixed prompts monthly across three engines and record four things: whether an AI answer appeared, whether you were named, whether your domain was cited, and which sources appeared instead. That is free and it beats most dashboards.
Where to go next
Start with the free half. Open your robots.txt and search it for OAI-SearchBot, PerplexityBot and Googlebot. That takes two minutes and it is the check most likely to be wrong.
Then run thirty of your buyers’ questions and write down which sources appear instead of you. Those two steps cost an afternoon between them. They will tell you which of the four layers is your real constraint.
Run the technical audit · Write for citation · Book a growth audit
Sources and further reading
- Ahrefs. Study of 1.4 million ChatGPT prompts, roughly 23.4 million cited and 23.4 million non-cited URLs. 49.98 percent of retrieved URLs cited. Title semantic similarity 0.602 cited against 0.484 non-cited. Natural-language slugs cited 89.78 percent against 81.11 percent. Citation rate by channel: search 88.46 percent, news 12.01 percent, Reddit 1.93 percent, YouTube 0.51 percent, academia 0.40 percent. The authors warn that non-cited URLs are 67.8 percent Reddit, so per-channel figures are more reliable than aggregate comparisons. Link
- Ahrefs. Controlled study of 1,885 pages adding JSON-LD schema between August 2025 and March 2026, matched against roughly 4,000 control pages, using difference-in-differences over 30-day windows. Google AI Overviews minus 4.6 percent and statistically significant, AI Mode plus 2.4 percent and ChatGPT plus 2.2 percent, both indistinguishable from zero. Every page already had 100 or more AI Overview citations, so it tests already-cited pages rather than unknown ones.
- Ahrefs. Study of 137,210 domains with traffic in May 2026, measured in server logs. 28 percent published an llms.txt, and 97 percent of those received zero requests. Adoption should be read as an upper bound, since the sample skews technical. Link
- Muck Rack. “What Is AI Reading?” May 2026. More than 25 million cited links across ChatGPT, Claude and Gemini. 84 percent of citations from earned media, 0.3 percent from paid. Link
- BrandMentions, 21 September 2026. Approximately 410,000 brand mentions across 240 brands in six sectors, 90 days to 17 July 2026. Repeatedly visible brands drew 70 to 80 percent of their mention footprint from third parties, against 30 to 40 percent for rarely visible brands, and 50 to 60 percent of those mentions carried no link. Correlation is the ceiling of what the data supports. Link
- Seer Interactive. Content recency study across 7,683 dated pages and 47,097 citations, March to June 2026: 75 percent of cited pages updated within a year against 42 percent recently published. Separately, 2025 click-through averages of 3.35 percent with no AI Overview, 2.07 percent when cited and 0.94 percent when not. Link
- OpenAI, Perplexity and Google crawler documentation, read September 2026. OAI-SearchBot surfaces sites in ChatGPT search while GPTBot crawls for model training. Google states Google-Extended does not impact inclusion in Google Search.
- Google Search Central. AI search optimisation guidance, 15 May 2026. Confirms AI Overviews and AI Mode run on core Search ranking and quality systems with no separate AI index, and warns against providers guaranteeing rankings.
- Deliberately excluded, and named so you can check: the 68 percent zero-click figure, 13x ROI claims, GEO case studies claiming 4,900 percent revenue growth or $90 million pipeline, SERP pixel-share percentages, vendor agent accuracy rates, and the claim that 44.2 percent of citations come from the first 30 percent of a page. None traces to a published method we could open.
Latest Blogs
Most advice on this is inference dressed as fact. OpenAI documents exactly one thing clearly, which is access. Here is what is documented, what is inferred, and how to verify the rest for yourself.
Being named in an AI answer and having your page used as its source are different outcomes with different causes and different fixes. The gap between them is the most useful diagnostic in AI search, and most reporting collapses it into one number.
Every platform reports share of voice and almost none of them define it the same way. Here is the definition worth using, the three choices that change the number, and how to measure it yourself without buying anything.
Get your hands on the latest news!
