Optimise Content for AI Search

To optimise content for AI search, make the relevant answer accessible, clear and supported by evidence. Check what each platform actually reports before deciding whether a change worked. The practical work overlaps with ordinary SEO and useful writing; a heading pattern or schema block cannot guarantee a citation.

My historical Bing Copilot portfolio case study reports 120,409 citations across 11 sites and a 57:1 citation-to-organic-click ratio. The published case study does not give enough reporting-window detail to verify that ratio as a matched comparison. I would not use it to predict referral traffic or describe today's performance.

How GEO, AEO and SEO Overlap

Generative Engine Optimisation (GEO) describes work to improve discovery and accurate use of a business or its content in generated answers. Answer Engine Optimisation (AEO) often describes answering questions in search features and assistants. SEO covers discoverability and performance in search more broadly. The labels overlap and are not separate, fully specified ranking systems.

A page's ranking, URL citations and brand recommendations can differ. Start with the question a reader needs answered and the evidence available, rather than assigning a different writing recipe to each label.

A Ranking Position Does Not Guarantee a Citation

A ranking report cannot tell you what appeared in a particular generated answer. Inspect the answer itself and distinguish a URL citation from a mention or recommendation. If you quote a citation study, retain its platform, prompt sample, dates and metric; a correlation in one dataset does not reveal a universal selection rule.

I have observed reference and troubleshooting pages among the cited pages in my portfolio. That is an editorial observation, not a controlled test of structure. Topic, source demand, visibility and concurrent changes could also explain the difference.

Check the Platform Before Prescribing a Fix

For Google, use its current AI search guidance. It calls for normal SEO and useful content, without special AI schema or a required chunk size. That is Google's guidance, rather than evidence of identical behaviour across every assistant.

For ChatGPT search, OpenAI documents OAI-SearchBot separately from its training crawler GPTBot and user-triggered fetcher ChatGPT-User. Inspect access for the relevant role. Training permission does not guarantee search selection.

For Perplexity, its crawler documentation distinguishes search access from user-requested fetching. It does not establish a public formula for which retrieved page receives a citation. Check actual answers instead of prescribing a reverse-engineered ranking recipe as fact.

Keep observations dated

An answer sample describes the product, mode and dates tested. Selection can vary between prompts and over time. Keep the original captures and record unavailable runs as missing; do not turn a change in a third-party sample into a confirmed explanation of a platform's internal behaviour.

Keep Answer, Evidence and Context Together

Read a section as if someone encountered it without the rest of the page. Can they identify its subject, understand the answer, check the source and see the main limitation? Repair what is missing without forcing every section to the same length. Keep a process with its prerequisites and a comparison with its criteria.

Here is a real correction from my portfolio reporting. The earlier wording left a reader with the wrong conclusion about the health affiliate site's AI referrals.

Before, from my earlier article:

AI referral traffic did not replace the lost Google traffic in volume, but it did not disappear when Google visibility did.

That sentence left out the site, dates and later decline. A 90-day total could contain both a period of stability and a fall.

After, using the later check:

On my health affiliate site, Google impressions fell from 20 July 2026. Recorded AI-assistant sessions held until late July, then fell from the week starting 3 August. The weekly average was 127 across the complete weeks starting 1 June through 27 July, compared with 28 across the weeks starting 10 August through 21 September.

My case study uses GSC and GA4 reports pulled on 4 October 2026. Its GA4 weeks use the property's America/Los_Angeles timezone and an eight-source assistant definition. The later decline overlapped my 2 and 3 August site changes; the data does not establish whether those changes or Google's decline caused it.

The repair makes the finding easier to understand and quote accurately. It is not evidence of a citation lift. My observed association between clear reference content and portfolio citations remains a correlation, with no isolated structure effect established.

Does Schema Markup Help?

Use standard structured data when it describes the visible page accurately. Keep business identity and genuine profile links consistent. A schema validation pass does not establish that an assistant will cite or recommend you.

Google's guidance does not require structured data for its AI search features. Its 8 May 2026 documentation update also confirms that FAQ rich results stopped appearing from 7 May. This does not establish an alternative FAQPage citation benefit. Add FAQs when they help readers, and use applicable markup for its documented purpose. My guide on adding schema markup covers implementation.

Does llms.txt Help?

An llms.txt file is a proposed guide to a site's content, rather than a search permission file like robots.txt. Google says it ignores these files for Search visibility. For a different service, check its documented support and intended use. I would not promise a citation improvement from adding the file, or make it a priority over correcting inaccessible or inaccurate pages.

How I Measure AI Search Performance

Keep three reporting sources separate:

1. Bing Webmaster Tools: AI Performance reports citations across supported AI surfaces. Page counts describe citation activity; sampled grounding queries are retrieval phrases, not a complete buyer-prompt list or a query-to-page attribution table.
2. Google Search Console: the Generative AI report reports impressions for AI Overviews and AI Mode. Google's help page dates worldwide rollout to 31 August 2026. Property and page aggregation differ, and exported zeros can represent unavailable values. It does not provide an AI-click total.
3. GA4: identifiable assistant session sources record attributed visits. Missing referral information leaves an unmeasured attribution gap; do not assign Direct sessions to AI without evidence. My AI referral traffic study explains why exposure and recorded visits need separate treatment.

Add sampled answer captures and actual enquiry records. A citation can support another brand's recommendation, and a form event can be a test or duplicate. None of these reports alone establishes qualified leads, revenue or the effect of a content edit.

What To Do With an Existing Page

Start with a valuable existing page. Check access and the answer a reader needs, repair unsupported claims, and keep the evidence and limitations connected. Save the original and revised section, the source date and the actual publication date. After an observed recrawl where available, compare later results with the baseline and record other changes.

I cover these checks in a technical SEO audit and dedicated AI search optimisation work. Get in touch if you want a second pair of eyes on the evidence and next step.

Related Articles

• What is LLM Optimisation? - the broader discipline behind AI content visibility
• AI Referral Traffic Study - recorded assistant visits and their measurement limits
• What is Entity SEO? - keeping business identity verifiable
• How to Add Schema Markup - implementation steps for any platform
• 120K Bing Copilot Citations: Portfolio Case Study - the historical citation figures, with matched-window verification still needed