Guide · 2027 edition

AEO best practices for 2027

What to keep, what to start and what to stop doing to get cited by AI assistants next year, based on what Google, OpenAI, Microsoft and Cloudflare changed in 2026 and on our own measurements.

The AEO best practices for 2027 are: set separate CDN rules for search, agent and training bots; cover a buyer's follow-up questions on one strong page; open every section with a self-contained answer; publish first-hand data; make your company unambiguous; earn mentions on sites models already trust; keep money pages current; make pages readable by agents; and measure with first-party reports plus repeated runs.

Google said in July 2026 that optimising for its AI features is still SEO, and for Google that is right. ChatGPT, Perplexity and Copilot run their own retrieval with their own crawlers and their own idea of who is worth naming. AEO in 2027 is SEO plus the parts Google’s guide does not cover. That is the gap this list closes.

What changed in 2026 that shapes 2027?

Five changes reset the baseline:

  • Google published its own AI optimisation guide (updated 10 July 2026). It says AI Overviews and AI Mode run on core Search ranking. It names query fan-out, where the model splits one question into several related searches, and says llms.txt, content chunking and AI-specific rewriting are not needed.
  • First-party AI reporting arrived. Bing Webmaster Tools launched an AI Performance report in public preview in February 2026, showing citations and the grounding queries behind them across Copilot and AI summaries. Google added generative AI performance reports to Search Console in June 2026.
  • The CDN became the real gatekeeper. Cloudflare now sorts bots into Search, Agent and Training. From 15 September 2026, new domains block Training and Agent bots by default on pages that carry ads.
  • OpenAI split its crawlers by job. OAI-SearchBot decides whether you can appear in ChatGPT search, GPTBot is for training, and ChatGPT-User fetches on a user’s behalf.
  • Agents started reading sites directly. Google’s guide notes browser agents read screenshots, the DOM and the accessibility tree, and names the Universal Commerce Protocol as an emerging standard.

1. Set separate bot rules for search, agents and training

A single “block AI bots” switch is no longer safe. Under Cloudflare’s new model, crawlers that do both search and training follow the strictest rule. Blocking Training therefore also blocks Googlebot, Applebot and Bingbot, even if Search is allowed. On the OpenAI side, blocking OAI-SearchBot removes you from ChatGPT search answers. Blocking GPTBot only opts you out of training.

Do this: allow Search and Agent bots on every page a buyer might be sent to. Decide on Training as a separate business decision. Then check the status code each named agent actually receives, because the CDN can say no while robots.txt says yes. Training vs retrieval, in full →

2. Cover the follow-up questions on one strong page

Query fan-out means a model answering “best contract software for in-house legal teams” also runs searches for pricing, integrations and alternatives. The page that answers the main question and its likely follow-ups gets retrieved on more of those searches.

Do this: for each high-intent page, list the five or six questions a buyer asks next, and answer them on the same page under their own headings. Do not create a separate page for every wording. Google’s guide treats pages made mainly to manipulate AI responses as scaled content abuse.

3. Open every section with a self-contained answer

Google says you do not need to write specially for AI or break content into tiny pieces. It does ask for clearly organised pages with headings and sections, and that is what this practice is. When we checked 7,663 headings across 50 funded UK SaaS sites in October 2026, 13% were phrased as a question a buyer would ask. Only 27% were followed by a direct answer of 25 to 90 words.

Do this: make the heading the buyer’s question. Answer it in the first sentence or two in a way that still makes sense when quoted on its own. Put the detail underneath. How to write content AI will quote →

4. Publish something only you can publish

Google’s guide asks for a “unique point of view” and non-commodity content, and calls generic listicles low-insight. An assistant has no reason to cite a page that says what twenty other pages already say. It does have a reason to cite the only page with a number.

Do this: turn something you already hold into a citable asset: product usage data, benchmark results, pricing transparency, a survey of your customers. Publish it with the method stated and the date on it. An example: our 2026 UK SaaS citation study →

5. Make it impossible to confuse your company with anything else

Models confirm what a company is partly by matching its site against outside profiles. In our October study, 26 of 50 funded UK SaaS sites had no sameAs links to those profiles in their Organization markup, and 35 had two or fewer.

Do this: publish Organization schema with one description, identical everywhere it appears. Link sameAs to every real profile (LinkedIn, G2, Capterra, Crunchbase). Add a Person node for the founder or named authors. If your name collides with anything, add an explicit disambiguating description. Keep schema matching the visible page; Google’s structured data rules require it. Entity clarity, a worked case →

6. Earn mentions on the sites models already trust

Ahrefs studied 75,000 brands across ChatGPT, AI Mode and AI Overviews. Branded web mentions correlated with AI visibility at about 0.66 to 0.71, and YouTube mentions at about 0.74. Domain Rating came in at 0.27 to 0.33. That is correlation, not cause, but the direction is consistent: models name brands that other credible pages name. Google adds the guardrail: inauthentic mentions do not help, because its spam systems filter them.

Do this: list the roundups, review sites, comparison pages and trade publications that appear when you ask an assistant your category question. Earn a place on them through outreach, reviews from real customers, and data journalists want to use. Never buy placements or post fake reviews.

7. Keep the pages that make money current

Ahrefs analysed about 17 million AI citations and found cited pages were on average 25.7% fresher than pages ranking in classic Google results. The effect was strongest in ChatGPT and weakest in AI Overviews. The average cited page was still almost three years old, so freshness is one factor, not the whole game.

Do this: put your money pages (pricing, comparisons, “best X for Y” answers) on a quarterly review cycle. Update facts that changed and show the real “last updated” date. Do not change dates without changing content. Google has warned against it, and it erodes trust.

8. Build pages an agent can read and use

In 2027 more visits will come from agents comparing vendors for a person rather than from the person directly. Google’s guide says browser agents read screenshots, the DOM and the accessibility tree, and recommends semantic HTML.

Do this: put core facts (pricing, plans, integrations, compliance, contact routes) in plain HTML text, not images, PDFs or content that only appears after a click. Use real headings, lists, tables and labelled form fields. A page that works with a screen reader generally works for an agent.

9. Measure with first-party reports plus repeated runs

Two sources of truth now exist that did not a year ago: Search Console’s generative AI reports and Bing’s AI Performance report, with grounding queries and per-page citations. Neither covers ChatGPT or Perplexity, and assistant output varies between runs. A single screenshot proves nothing.

Do this: connect both webmaster tools. Track AI referral traffic in GA4. Freeze a set of real buyer questions and run each one at least three times per engine, recording the spread. Do not change the question set when results are inconvenient. Our measurement protocol → · AI traffic in GA4 →

10. Stop doing what the evidence does not support

  • llms.txt as a citation lever. Google ignores it. In our October study, sites with one were named 0.95 times on average against 1.14 for sites without. It is cheap and harmless, and some agent tools read it, so keep it. Just do not expect it to move citations. What llms.txt actually does →
  • A page for every prompt variation. Models understand synonyms. Thin variant pages count as scaled content abuse.
  • Treating a technical score as the goal. In our study, technical readiness correlated with being named at only 0.24. The best-built site, scoring 95 out of 100, was not named once. Access is necessary, not sufficient.
  • Bought or manufactured mentions. They are filtered out, and they put the whole domain at risk.

2027 at a glance

PracticeFor 2027Why
Separate CDN rules for search, agent and training botsStartCloudflare defaults changed 15 Sep 2026; strictest rule wins
Follow-up questions covered on one pageStartQuery fan-out, named in Google’s July 2026 guide
Semantic, agent-readable pagesStartAgents read the DOM and accessibility tree
Search Console and Bing AI reportsStartFirst-party citation data, new in 2026
Answer-first sections under question headingsKeepOnly 27% of headings we checked had a direct answer
Organization and Person schema with sameAsKeep26 of 50 sites we checked had no sameAs links
Earned third-party mentionsKeepStrongest correlate in Ahrefs’ 75k-brand study
Quarterly refresh of money pagesKeepAI-cited pages are 25.7% fresher (Ahrefs)
Repeated runs on a frozen question setKeepAssistant output is non-deterministic
llms.txt as a ranking leverStopIgnored by Google; no lift in our data
Pages per prompt variationStopScaled content abuse
Bought mentions, date bumps without changesStopFiltered as spam; erodes trust

How this page is kept current

We review this list every quarter, against the sources below and against fresh runs of our own question sets. When a platform changes its rules, the practice and its date change here. The 2028 edition will replace this one at the same address.

Sources

Every figure on this page links to the page that published it. Correlations are labelled as correlations. Our own figures come from one engine on one day and are described as such in the study.

FAQ

Related questions

For Google, no. Its July 2026 guide says optimising for AI Overviews and AI Mode is SEO. ChatGPT, Perplexity and Copilot run their own retrieval with their own crawlers, so bot access, entity clarity and measurement still need work Google’s guide does not cover.

Keep it if you have one. It is cheap, and some agent tools read it. Google ignores it, though, and in our October 2026 study sites with llms.txt were not named more often than sites without. Do not treat it as a citation lever.

Getting named on the third-party pages models already trust for your category: roundups, review sites and trade press. In Ahrefs’ study of 75,000 brands, branded web mentions correlated more strongly with AI visibility than any on-site factor they measured.

Connect Search Console’s generative AI reports and Bing Webmaster Tools’ AI Performance report for first-party data. Track AI referral traffic in GA4. For ChatGPT and Perplexity, run a frozen set of buyer questions at least three times per engine and record the spread.

Yes. From 15 September 2026, new Cloudflare domains block Training and Agent bots by default on pages that carry ads. Blocking Training also blocks crawlers like Googlebot that do both search and training. Check what each named bot actually receives, not just robots.txt.

We review it every quarter against the sources listed and fresh runs of our own question sets. The 2028 edition will replace it at the same address.

See how your site measures up against the 2027 list.

Send your domain. The free check tests bot access, extraction and entity signals, then asks AI engines the questions your buyers ask.