Three technical tactics are being sold as AI search requirements right now: a special schema type for AI, breaking your content into very small chunks, and publishing an llms.txt file.
Official guidance contradicts all three. That does not make them worthless, but it does mean the confident claims attached to them are running well ahead of the evidence, and budget spent on them is usually budget not spent on the things that demonstrably matter.
Google says no special AI schema is required, that there is no requirement to break content into tiny pieces, and that it currently ignores llms.txt. Standard accurate structured data, genuinely readable content and crawler access are the technical foundations that hold up. Treat everything beyond that as a hypothesis, not a requirement.
This page separates two things deliberately. Documented behaviour is what search and model providers state in their own guidance, and it is labelled as such throughout. Reasonable practice is what we recommend based on how retrieval works, and it is labelled as interpretation. We have not run controlled citation experiments at a scale that would support causal claims, so we do not make any. Anyone telling you a single technical tactic guarantees citations is overstating what is currently known.
Does schema markup help you get cited by AI?
Standard structured data helps machines understand your page accurately. There is no special AI schema type, and Google has stated that no AI-specific markup is required.
That distinction matters commercially, because vendors have started selling "AI schema" as a discrete product. What actually earns its keep is ordinary schema, implemented correctly, describing things that are genuinely on the page.
| Claim | Status | What to do |
|---|---|---|
| A special AI schema type is required | Contradicted by Google guidance | Use standard schema types accurately |
| Schema guarantees AI citation | No evidence supports a guarantee | Treat schema as clarity, not leverage |
| Accurate structured data aids machine understanding | Well supported | Implement and validate properly |
| Schema must match visible page content | Google requirement | Never mark up content users cannot see |
The types worth prioritising are the boring ones. Organization establishes who you are and ties your brand entity together. LocalBusiness carries the location and contact facts that listings also need to agree with. FAQPage structures genuine questions and answers, but only where those questions actually appear on the page. Article and Author support attribution and expertise signals. Product and Offer carry pricing and availability for ecommerce.
The failure mode is almost never a missing exotic type. It is a mis-implemented common one: Organization markup that disagrees with the footer, FAQPage markup for questions not visible on the page, or LocalBusiness data that contradicts the Google Business Profile.
Our view is that schema helps AI citation indirectly, by making your entity unambiguous rather than by acting as a ranking signal. A model that can confidently resolve who you are, what you do and where you operate has fewer reasons to reach for a competitor whose facts are clearer. We consider that reasoning sound, but it is interpretation, and we are labelling it as such.
Do you need to break content into small chunks for AI?
No. Google Search Central has stated plainly that there is no requirement to break your content into tiny pieces for AI to better understand it.
The chunking advice circulating in the AEO market comes from a real technical concept applied too literally. Retrieval systems do segment documents when building an index. That is a fact about how retrieval works. It does not follow that authors should pre-chunk their prose into fragments, and writing that way usually produces worse content for humans without any documented citation benefit.
| Common claim | What guidance says | Practical takeaway |
|---|---|---|
| Paragraphs must be under N words | No such requirement stated | Write clear paragraphs of natural length |
| Every section needs a 60-word answer block | Vendor claim, not official guidance | Answer directly, but do not pad to a word count |
| Long-form content cannot be cited | Not supported | Depth is fine when it is well structured |
| Structure helps retrieval | Consistent with how retrieval works | Use real headings and logical order |
What does hold up is structure. Descriptive headings, a direct answer near the top of each section, and a logical document order all make content easier to parse, for models and readers alike. That is different from chopping your writing into uniform fragments.
The practical version: put the answer first under each heading, then expand. Not because 60 words is a magic number, but because a reader who wants a quick answer gets it, and a model extracting a passage finds a clean one. Our guide to structuring content for AI search goes deeper on the writing side.
Is llms.txt worth publishing?
Google has stated it currently ignores llms.txt. It is not a Google ranking or citation lever, and any vendor presenting it as one is not describing current documented behaviour.
The proposal itself is reasonable in principle: a plain-text file listing your most useful content for language models, comparable in spirit to robots.txt or a sitemap. The problem is adoption. A file that major consumers ignore does not do the job it was proposed to do, however sensible the idea.
| Question | Answer |
|---|---|
| Does Google use it? | Google has stated it currently ignores llms.txt |
| Is it an industry standard? | A proposal, not a ratified standard |
| Does it cost much to publish? | Very little, it is a static text file |
| Should it be a priority? | No, well behind crawler access and accurate schema |
| Could that change? | Yes, which is why it is worth revisiting periodically |
Our position: publish it if you like, because it is cheap and harmless, but do not let it displace work that is documented to matter. If an AEO proposal leads with llms.txt, ask what the rest of the technical plan looks like.
What actually matters technically
If the three headline tactics are weaker than advertised, what should the technical budget buy? These are the items with the clearest support.
Yext's analysis of 6.8 million AI citations found 86% came from brand-managed sources, split 44% first-party websites and 42% business listings (Yext, Oct 2025). That is the single most useful technical prioritisation signal available: the surfaces you control account for most of what gets cited.
Crawler access comes first. If AI crawlers cannot fetch your pages, nothing downstream matters. OpenAI documents OAI-SearchBot as the crawler governing ChatGPT search discovery, and blocking it removes you from that surface entirely. Check your robots.txt, your CDN rules and your bot-management settings, because AI crawlers are frequently blocked by default in security configurations without anyone deciding to block them.
Rendering comes second. Content that only exists after client-side JavaScript execution is at higher risk of being missed. Server-rendered content is the safer default for anything you need cited.
Accurate, validated schema comes third. Standard types, matching visible content, validated rather than assumed.
Entity consistency comes fourth. Your name, address, category and service descriptions should agree across your site, your schema and your listings. This is where technical work meets the 42% listings share, and it is usually the cheapest available gap to close.
Confirm your robots.txt and CDN are not blocking AI crawlers. This single misconfiguration removes a site from ChatGPT search discovery entirely, and it is common because bot-blocking rules are often inherited from security defaults rather than chosen deliberately. It is worth checking before any other technical AEO work.
A technical checklist you can work through
Run these in order. Each one is verifiable, which is the point.
- Confirm AI crawler access. Check robots.txt, CDN and WAF rules for blocks on AI crawlers including
OAI-SearchBot. - Verify server-side rendering for the pages you most want cited.
- Validate your structured data rather than assuming it works. Fix errors before adding new types.
- Reconcile entity facts across site, schema and every business listing you control.
- Audit your answer structure. Does each key page answer its question in the opening lines, or does it open with positioning?
- Check page-level basics. Titles, headings and internal links should describe content accurately.
- Establish a measurement baseline before making changes, so you can tell whether anything moved.
Steps 4 and 5 usually produce the most movement, and neither is exotic. For the off-site half of the work, see our guide to earning AI citations through digital PR, reviews and third-party mentions. For the entity layer, entity SEO for AI search covers how models resolve brand identity.
For where this technical layer sits relative to the wider discipline, see AEO vs GEO vs SEO.
What we would need to prove causation
Being straight about the limits here matters, because the market is full of confident causal claims that no published evidence supports.
To say "changing X caused citation Y", you would need a fixed prompt set, repeat runs across engines to establish baseline variance, a controlled change on one variable, a comparison group of unchanged pages, and enough repetition afterwards to distinguish a real shift from normal model volatility. Very little published work in this category meets that bar.
Until it does, the honest framing is: crawler access, accurate structured data, entity consistency and answer-first content are well-founded practice. Specific thresholds, magic word counts and proprietary schema products are hypotheses being sold as requirements.
Want to know which technical layer is actually blocking you?
Book a free AI Visibility Audit. We check crawler access, structured data, entity consistency and answer structure, then show you where you appear across ChatGPT, Perplexity, Gemini and AI Overviews.
Book Your AI Visibility AuditFrequently Asked Questions
Does adding FAQ schema get me into AI answers?
Not on its own. FAQ schema helps machines parse questions and answers you have genuinely published, which is useful, but it is not a citation lever and Google does not require any AI-specific markup. It also has a hard requirement: the questions and answers must be visible on the page. Marking up questions that users cannot see is a structured data violation, not a shortcut.
If Google ignores llms.txt, why do so many agencies recommend it?
Partly because the idea is genuinely sensible, and partly because it is a visible, cheap deliverable that looks like technical progress on a report. It costs almost nothing to publish, so there is no strong reason to refuse. There is a strong reason not to let it substitute for crawler access checks, schema validation and entity work, which are documented to matter.
How long should paragraphs be for AI search?
There is no documented requirement, and Google has explicitly said content does not need breaking into tiny pieces. Write paragraphs at whatever length communicates the idea clearly. The structural habit that genuinely helps is answering the question directly under each heading before expanding, which serves both a reader skimming and a model extracting a passage.
Should I block AI crawlers to protect my content?
That is a legitimate business decision with a direct consequence: blocking a crawler removes you from the surface it feeds. Blocking OAI-SearchBot removes you from ChatGPT search discovery. Some publishers make that trade deliberately for licensing reasons. What you should avoid is blocking accidentally through inherited security defaults, which is the far more common situation and costs visibility nobody chose to give up.
What is the single highest-value technical fix?
For most sites, reconciling entity facts across the website, schema and business listings. With 42% of AI citations coming from business listings (Yext), inconsistency there is expensive, and it is usually cheaper to fix than any content project. Crawler access is more urgent when it is broken, but it is broken far less often than entity data is inconsistent.
Ashur Homa
Built and scaled a digital brand to $100M+ in sales with zero ad spend. Has helped businesses generate millions through AI go-to-market strategy. Leads growth at Omni Eclipse.
Connect on LinkedIn
