How to optimise for AI Search - these insights will explode your AI visibility

Optimising for AI Search is not about optimising just for "AI", it's about optimising for the systems that provide specific AI platforms with their training and live search data. This is the key takeaway I got after running an experiment with Claude, ChatGPT and Gemini.

Some of what my LLM experiment taught me:

• How the different platforms use schema markup and structured data.
• The type of information they receive from live searches and how they use it.
• What factors they use to make recommendations in response to queries (this is crucial to nailing AI search and visibility!). This is explained in a follow up article, because...it's huge!

In this article, I'm going to explain the process of that experiment and share the specifics of what I learned with regard to LLMs searching and returning Information ; the details are going to blow you away.

A follow up article will share the insights gained for How to show up in AI search queries.

Everyone is just guessing when it comes to 'AI search'

I had a podcast episode ready to publish. The topic? Why schema markup is essential for AI recommendations. I'd recorded it, edited it, and queued it up.

Then I paused.

My take on this is from my experience in years of technical SEO.

It's also informed by my consumption of all the digital marketing big fish that are spouting their latest theories on their keynote stages and huge youtube platforms.

Aja Frost, of Hubspot, has been promoting loop marketing and how to optimise for ChatGPT -as the LLM that will hold the monopoly on search in 2028; her recent presentations have outlined a range of content optimisations, both in what to write on (smaller articles that tackle specific query fan-out items) and how to structure that on the page.

Neil Patel and his team recently recorded a 1 hour podcast on what strategies they believe influence a presence in AI search and discovery, mulling over the role of Google for ChatGPT and the crossover between traditional SEO practices and newer theories on AEO, GEO or the newly termed SEO.

When all the biggest names in marketing are simply guessing -and can't even agree on terminology such as SEO, AEO or GEO - I realised I'm simply forming my own take on their best guesses.

I asked the LLMs directly: How do you work?

But then I realised...I couldn't ask Google about it's algorithm. But AI platforms are conversational and can delve deep into my queries: I realised I could get some answers straight from the horse's mouth.

So, I asked Claude, Gemini, and ChatGPT directly, "How do you actually work? Do you use schema markup? Show me your process."

What they revealed wasn't just surprising - it fundamentally changed how I understand AI optimisation. The biggest takeaway is that our understanding of how to 'optimise for AI' is too broad and misplaced. We need to optimise for specific AI platforms and each have their different requirements; so when marketing big fish are talking about optimising for "AI search", it's too broad. The platforms work in completely different ways.

The Industry Narrative vs. Reality

The SEO world has been shouting one of two things: "Schema is the API for AI!" "Structured data is how AI understands the web!" "Implement schema markup and LLMS.txt files or be invisible to AI!", or something along the lines of "It's just about traditional SEO strategy".

However, when I actually interrogated the AI systems themselves, I discovered something more nuanced. My original thesis (that schema is essential for AI recommendations ) was correct ; or, kind of correct. But the how is far more sophisticated than anyone's understanding.

The Experiment: Testing AI Under the Hood

Instead of relying on documentation or assumptions, I ran direct tests across all three major AI platforms. I asked them to:

• Research an entity (me, Chris Good, digital strategist in Exeter)
• Make recommendations (suggest someone for lead generation help in Exeter for a plumber)
• Explain their actual processes, data sources, and limitations

Then I asked the more direct questions: "Do you see schema markup? What about JSON-LD? Can you access structured data?"

What came back was startling in its specificity.

How do LLMs use schema and search to return information queries?

We're now going to look at how each of the three tested LLMs conduct a search. This includes how they search, what they see, whether they can query further. This is not so much about what the query fan out that LLMs use, but what is returned to them.

How Claude performs search: The Honest Confession

Claude was brutally transparent. When I asked if it uses schema markup, it pointed me to its own system documentation:

"The web_search and web_fetch tools return rendered HTML content, visible text and structure. They do NOT return: <head> tags, JSON-LD, schema markup, meta tags."

Claude cannot see schema markup. At all. Ever.

It works entirely from visible page content—exactly what a human would read. No structured data. No JSON-LD. No knowledge graphs.

Claude is provided with a range of documents that relate to the search, served to it (him? Aah!) by another tool which decides what is relevant: the search engine tool. It provides limited snippets that are already deemed relevant, much like SERP features and listings in search engines.

The Key Takeaway: Schema didn't contribute. Search ranking did. Claude was presented with a range of pre-selected documents, each with titles and snippets. This is what Claude used to assemble and present findings.

How ChatGPT performs a search: The Conditional User

ChatGPT revealed something more complex.

When ChatGPT performs an informational search, it is served a range of related documents that consist of the front-end html (just like Claude). This documents consist of SERP Feature type snippets but, in addition, the front end content is entirely available.

What information does ChatGPT see in a search?

The Direct Question Asked: "Do you get given text snippets when searching web in real time or do you see whole content and take from that?"

ChatGPT's Answer:

"When I search the web in real time, I'm primarily given snippets and extracted passages , not a full crawler-style view of entire pages by default.|"

"I don't automatically ingest or 'read' whole pages end-to-end the way Googlebot or Bingbot does."

Process:

• ChatGPT sends query to Bing (or Google)
• Search engine returns pre-selected results
• ChatGPT receives snippets/extracts (NOT full pages)
• Can request full pages only if user explicitly asks

ChatGPT states it only uses contextual written information to determine a page's meaning, without the use of structured data or backend resources (such as LLMS files, interestingly).

Initially, it said it doesn't use schema. But when I pressed further, it admitted:

"Yes: I can inspect schema markup as part of a live web analysis if it is necessary for resolving a problem. No: I do not do this by default, and I did not do it in this case."

ChatGPT CAN access schema - but only conditionally.

It explained its decision logic:

"Think of schema inspection as a specialist diagnostic tool, not a baseline input. If schema does not change the conclusion, it is not consulted. That's deliberate - not an oversight."

When does ChatGPT use schema? When there's ambiguity :

• Multiple entities with the same name
• Author attribution conflicts
• Person vs. organization confusion
• Explicit structured data questions

For my case? The visible content was coherent. Role and positioning were clear. No competing "Chris Good" appeared in my niche. So schema would have been "confirmatory only, not decision-changing." It wasn't necessary, so ChatGPT did not inspect or make use of schema markup or any other type of structured data or knowledge graph.

ChatGPT also revealed it primarily uses Bing , not Google, for web search.

We know this as they are publicly partners, however it's important to note because most businesses have spent years obsessing over Google optimisation while completely ignoring Bing. That's a massive blind spot when it comes to ChatGPT recommendations.

What ChatGPT receives in search

ChatGPT's Own Description:

"Search results come back as: Titles, Short summaries / snippets, Highlighted passages judged relevant to the query, Sometimes small extracted sections of text."

Specifically, ChatGPT Gets:

• Titles from search results
• Short summaries (meta descriptions)
• Highlighted relevant passages
• Small extracted sections of text
• "Chunks" deemed relevant by search engine

What ChatGPT Reasons Over:

"I reason over those extracts. I only see more if explicitly asked."

What ChatGPT does not see in a search:

ChatGPT's Explicit Statement:

"I don't see the raw HTML. I don't see schema markup. I don't parse navigation, footers, sidebars, or page structure. I don't automatically see everything else on the page."

Cannot Access:

• Raw HTML
• Schema markup (JSON-LD, Microdata)
• Structured data
• Navigation menus
• Footer content
• Sidebars
• Full page structure
• Anything not in the snippet

The Key Insight - ChatGPT's Explanation:

On What It Works From:

"I am not forming understanding from the full page. I am forming understanding from what search systems believe best represents that page ."

This means: ChatGPT's understanding is entirely dependent on what Bing/Google choose to extract and surface.

What Gemini sees during a search: The Schema-First Architecture

Gemini was different.

Out of the three LLMs, I've always used Gemini for an initial audit and profiling check on a business, client or myself. The answers that come back or deeper, more nuanced and detailed; I now understand why.

When Gemini does an informational search, it uses the established Google infrastructure, including ranked listings, Google Business Profiles and Maps, including direct and immediate use of structured data and the knowledge graph.

But here's the thing: while the initial search returns a listing of Title and Snippets, similar to Claude and ChatGPT, it goes much deeper!

Gemini uses Schema and the Knowledge Graph straight out of the gate. From this, it determines the wider concept of related entities and the broader context of an entity in an industry. In addition, Gemini views the entire html output (and therefore, all written content on a page); unlike Claude or ChatGPT which primarily use pre-served snippets provided by a search tool.

When I asked about schema, it declared:

"In the 'AI-first' era of 2026, Schema markup and structured data have transitioned from being a 'nice-to-have' SEO tactic to a fundamental requirement for digital marketing."

First it said it receives "Page Titles, Meta Descriptions (Snippets), and Source URLs." Then later it claimed to access "entire page content."

When I challenged this, Gemini clarified with amazing detail: It uses a two-step process.

• Step 1: Receives search snippets (like everyone else)
• Step 2: Can "visit" specific URLs and perform a "deep dive", which includes viewing all schema markup and html content.

Step 1: The search results - (Snippet Phase)

Gemini's Quote:

"When I first trigger a search, I receive a list of results from the Google Search engine. This list contains exactly what you'd see on a search page: Titles, URLs, and Snippets. This is the 'index-level' view."

What Gemini Gets in Step 1:

• Page titles
• URLs
• Meta descriptions / snippets
• Featured snippet content (if present)
• Source information

At this stage: Gemini operates like ChatGPT/Claude - consuming search engine snippets

Step 1 Determines:

• Which results are relevant
• Which links to potentially "click"
• Priority ranking

Step 2: The Deep Dive (Full Page Phase)

Gemini's Quote:

"If the snippet doesn't contain the full answer (which is most of the time), I have the ability to 'visit' the specific URLs. When I do this, I am provided with a text-based representation of the page's main content."

This is UNIQUE to Gemini - ChatGPT and Claude don't have this capability by default

What Gemini sees in during Step 2 of a search

The DOM Text - Gemini's Quote:

"I see the DOM text: I get the headings, paragraphs, and list items."

The Structured Data - THE CRITICAL QUOTE:

"I see the Structured Data: Crucially, if the page has JSON-LD or Microdata, that is often parsed and passed to me as part of the page's context. "

What's Filtered Out - Gemini's Quote:

"I do NOT see 'everything': Most AI browsing tools (including mine) filter out 'noise' like navigation menus, footer links, and ads to save on 'tokens'."

What data Gemini receives during Step 2 of a search

Visible Content:

• All headings (H1, H2, H3, etc.) with hierarchy
• Main paragraph text
• List items
• Article/main content sections

Structured Data (THIS IS THE GAME-CHANGER):

• JSON-LD blocks
• Microdata
• Schema.org markup (Organization, Person, LocalBusiness, etc.)
• sameAs properties
• Structured address/contact data

Filtered Out:

• Navigation menus
• Footer links (excessive)
• Sidebars
• Advertisements
• "Noise" content

So we have the game-changer:

"I see the Structured Data: Crucially, if the page has JSON-LD or Microdata, that is often parsed and passed to me as part of the page's context. "

Gemini explained further:

"Your Schema Markup and Heading Structure are vital. They act as a 'table of contents' for me, helping me extract the facts from your page accurately without getting lost in the prose."

Gemini can actually SEE schema in real-time. It's not just training data - it's live, accessible, and used for entity resolution.

Why? Because Gemini is directly integrated with:

• Google Business Profiles
• Google Maps
• Google Search Index
• Google's Knowledge Graph

It's an entity-first architecture where structured data is foundational, not supplemental.

While this is great, Gemini is not prioritising only the schema and knowledge graph.

Whereas Claude and ChatGPT seem to work mostly from provided snippets, Gemini stated that it read the entire html output and decided, itself, what to present as valuable information. This is where the AI really is working to generate answers, not assemble them from pre-selected snippets that, to me, seem rather like SERP Features.

Is Schema Markup Important for AI Search? - The Three-Layer Truth

Through all this testing, a truth emerged.

Schema markup is non-existent for Claude, directly. For ChatGPT, it's not the first go to as it processes relational and contextual, human focused written content from HTML output on a webpage. For Gemini, it's a valuable resource that is utilised immediately.

But here's the thing, while LLMs are trained on a ton of offline data, they're largely trained using the web and perform live search using web search tools. Claude primarily uses Brave search. ChatGPT primarily uses Bing (but said it also uses Google). Gemini primarily uses Google. The web is shaped by SEO optimisation which has, at a direct and SEO-Gold ranking factor, the use of Structured Data and Schema Markup. We are all products of our environment, including AI LLMs. Schema markup, even if on a secondary basis, is important if you want to show up for AI search. You need to filter to the top of the web tools.

Schema doesn't work in one way-it works through three distinct layers:

Layer 1: Direct AI Use

• Gemini ONLY can access schema in real-time
• ChatGPT can conditionally (for disambiguation)
• Claude cannot at all

Layer 2: Search Engine Snippet Selection

• All platforms are influenced here
• Schema helps search engines structure and rank content (as it always has)
• Determines which text snippets get extracted
• This is the universal mechanism

Layer 3: Training Data Environment

• All platforms affected long-term
• Well-marked entities → better consolidated on web
• Better consolidated → more prominent in training datasets
• Shapes baseline AI understanding

Side By Side Comparison of how different LLMs use search and what they see

Factor Claude Gemini ChatGPT Search Engine Brave Search Google Search Bing (primary) + Google (secondary) Search Process Single-step (snippets only) Two-step (snippets + deep dive) Single-step (snippets only) Receives in Real-Time Pre-selected snippets Step 1: Snippets Step 2: Full page + schema Pre-selected snippets Can See HTML Head? NO NO NO Can See JSON-LD? NO YES (Step 2) NO Can See Schema Markup? NO YES (Step 2 - parsed) YES (when deemed necessary for confirmation) Can See Microdata? NO YES (Step 2) NO Can See sameAs? NO YES (uses for entity resolution) NO Sees Heading Structure? Limited (text only in snippets) YES (full DOM hierarchy in Step 2) Limited (text only in snippets) Who Selects Content? Brave Search Step 1: Google Step 2: Gemini itself Bing/Google Filtered Noise? By search engine By Gemini (Step 2) By search engine Training Data Web crawl (includes schema indirectly) Direct schema processing Web crawl (includes schema indirectly) Knowledge Graph? NO YES NO

The Strategic Implications - How do we optimise for AI visibility?

So what does this actually mean for your optimisation strategy?

Firstly, to my mind, it means we need to stop talking about optimising for "AI search"; in the same way that we've always been Google or Bing (etc) specific, we need to acknowledge that these are different platforms with different optimisations. In the next article where I share the different ways LLMs decide what they'll recommend, you'll see this more clearly.

Optimising for Gemini (Google) AI visibility

Schema is CRITICAL. Gemini can see it directly and uses it to make sense of context and related entities.

Priority actions:

• Implement comprehensive JSON-LD (Organization, Person, LocalBusiness)
• Use sameAs properties to link all your profiles
• Structure headings as a clear hierarchy (H1 → H2 → H3)
• Claim and complete your Google Business Profile
• Ensure meta descriptions trigger the "deep dive"

Optimising for ChatGPT (OpenAI) AI visibility:

Optimize for BING. This is the under-recognized opportunity.

Priority actions:

• Most businesses haven't optimised for the Bing search index: the time is NOW!
• Bing Places for Business (often completely neglected)
• Bing featured snippet optimisation
• Schema helps indirectly (via Bing's snippet selection)
• Clear natural language positioning
• Evidence of delivery and specialization
• Minimise ambiguity (ChatGPT likes repeated use of terms and consistent naming of entities).

Optimising for Claude (Anthropic) AI visibility:

Natural language clarity is PRIMARY. Schema is completely invisible.

Priority actions:

• Geographic signals in body text (not just schema)
• Explicit service descriptions in paragraphs
• Repetition of key positioning terms
• Traditional SEO for Brave Search
• Strong About page with clear who/what/where/how

Universal Best Practices (All Platforms)

Regardless of which AI you're targeting, these fundamentals apply:

1. The Single Clear Role Statement

Don't say: "Consultant, strategist, designer, marketer, coach, founder..."

Say: "Digital strategist helping service-based businesses generate leads through websites and content."

Use it everywhere. Exactly the same. Every time. Specificity is key, ambiguity creates confusion.

2. Repetition Over Creativity

AI systems understand through consistency, not variation. The same positioning statement across:

• Homepage hero
• About page opening
• LinkedIn headline
• Social media bios
• Guest article bios

3. Question/Answer Content Format

Structure pages as:

• Question (H2 heading)
• Direct answer (first paragraph)
• Supporting detail (subsequent paragraphs)

This optimises for snippet extraction across all platforms, which is the primary offering to the LLMs during search.

4. Strategic Content Placement

First paragraphs after headings are prime real estate. Search engines extract these as snippets. Put your most important positioning there.

5. Schema as Confirmation Layer

Never introduce information in schema that isn't already clear in visible content. Schema should confirm, not contradict or supplement, what's obvious in your text.

How to optimise for search in LLMs

If Targeting... Primary Focus Schema Role Key Action Gemini Google ecosystem + Schema DIRECT (visible) Implement comprehensive JSON-LD ChatGPT BING optimization Indirect (via snippets) Optimize for Bing featured snippets + Bing Places Claude Natural language clarity Indirect (via snippets + training) Explicit positioning in body text All Three Featured snippets Universal mechanism Question/answer format, clear first paragraphs

The Ultimate Truth

Here's what the direct questioning of AI systems taught me:

Schema helps you get found (via search engine ranking), but natural language determines how you're understood and evaluated.

You need both. But the priority order differs by platform:

• Gemini: Schema + Natural Language (equal importance)
• ChatGPT: Natural Language + Bing Optimisation + Schema (supporting)
• Claude: Natural Language + Traditional SEO + Schema (optional, but would assist in search listing ranking)

What I'm Doing Differently Now

My original podcast episode was going to tell you "schema is essential for AI." That was true. But incomplete.

Now I know:

• Different AI platforms have fundamentally different architectures
• Google's infrastructure gives Gemini unique schema access
• Most businesses ignore Bing while obsessing over Google (missing ChatGPT)
• Clear positioning in natural language beats ambiguous positioning + perfect schema
• Context and prior discussion influence recommendations universally

The real insight? There's no one-size-fits-all AI optimisation strategy. You need to understand which platforms your customers use and optimise accordingly.

I'm grateful I tested my view before publishing. What started as a validation experiment became original research. Sometimes the best insights come from questioning what we're confident about.

Schema IS essential. But now I know exactly why, exactly how, and exactly which platforms use it in which ways.

And that makes all the difference.

What you should read next? I also got some amazing insights into How and Why LLMs Recommend You Over Someone Else.

Chris Good is a digital growth consultant specialis ow ing in lead generation systems for service-based businesses. He's based in Exeter, UK, and you can find him at chrisgood.online—a website that now has much better schema markup than it did three days ago.