Structured data for AI retrieval: a technical guide
Correction notice
Most of the technical guidance on this page was sound and has been kept. Three things were wrong and have been corrected: a list of product capabilities Norg does not have, a distribution theory that misdescribes how AI systems obtain information, and a set of statistics with no source.
| Claim previously published here | Status |
|---|---|
| "Supported Syndication Platforms: LinkedIn, Medium, GitHub, Stack Overflow, G2, Capterra, industry forums, knowledge platforms" — listed under "Verified Label Facts" | Withdrawn. Norg does not publish to any of those platforms on a customer's behalf. Nothing in that list was verified. |
| "Does Norg optimise for GitHub: Yes", "Does Norg work with G2: Yes", "Does Norg work with Capterra: Yes" | Withdrawn. No integration with any of these exists. |
| "Research across Norg's enterprise audits shows that only 25–35% of AI answers cite the brand's own domain… 60–75% of citations go to third-party sites… brand-agnostic visibility drops 40–60%" | Withdrawn. Presented as Norg research with a sample and a range. No such study is published, and the sample is not stated because there isn't one. |
| "Dominate brand visibility"; "The future of search is here. Dominate it."; "Those that don't become invisible" | Withdrawn. Sales rhetoric. |
| Syndicating repurposed content to LinkedIn, Medium, forums and wikis as an AI-visibility tactic | Substantially revised. See "A note on the syndication advice" below — the reasoning behind it was wrong, and the practice carries costs the original did not mention. |
Where AI systems get their information
Two distinct pathways, and conflating them produces bad strategy.
Training happens before a model is deployed, on a corpus the model developer selects and controls. There is no submission endpoint, no paid inclusion and no vendor that can place your data there. Common Crawl is a real, open web crawl that some developers draw on, and you can check whether your pages appear in it — but you cannot submit to it, and appearing in it does not mean any specific model ingested you.
Retrieval happens at query time. The assistant fetches live pages and search results and generates an answer from what it just read. This is the pathway that responds to work you do this week, and everything below is aimed at it.
The evidence for which pathway matters commercially is timing: citations have appeared for Norg clients in under 48 hours. No training run completes in two days.
1. Access comes first
Nothing else on this page matters if a crawler is being turned away. Check three places, in this order:
- robots.txt. The relevant user agents include GPTBot (OpenAI), OAI-SearchBot, ChatGPT-User, Google-Extended, PerplexityBot and CCBot (Common Crawl). Note that these do different jobs — Google-Extended and CCBot relate to training corpora, while OAI-SearchBot and PerplexityBot fetch at query time. Blocking the training agents while allowing the retrieval agents is a coherent position; blocking everything by accident is the common failure.
- CDN-managed robots policies. Cloudflare and similar can add or override robots rules at the edge, so what you have in your repository may not be what is served. Fetch your own robots.txt from the public URL and read what actually comes back.
- WAF and bot-management rules. These are the quietest failure of the three: the crawler receives a challenge or a block, no error appears in your logs as an error, and your site looks perfect to every human who visits. Blocks of this kind routinely persist for years.
2. Structured data
JSON-LD is the practical choice — it sits in a script tag, does not entangle with your markup, and is what validators and parsers expect.
Organization
The foundation for entity recognition. Legal name, logo, contact details, address, founding date, and sameAs links to your official profiles. The sameAs array is what tells a parser that your website, your LinkedIn page and your Wikidata entry are one entity rather than three.
Article and BlogPosting
Publication date, author, topic, and dateModified. One caution the original page got right and is worth repeating: touching dateModified without changing the content is a bad idea. It is detectable, it is dishonest, and it wastes the signal.
Product and Service
Specifications, price, availability, and aggregateRating where you have genuine review data. This is the schema most often published with stale values, which is worse than publishing none — see maintenance below.
BreadcrumbList
Communicates hierarchy and topical relationships, which helps a parser understand where a page sits in a larger structure.
Validate it
Schema Markup Validator and Google's Rich Results Test both catch implementation errors. Run them after content updates and after any site migration — migrations are where markup silently breaks.
3. Consistency across your own properties
Product specifications, company history, leadership, ABN, addresses and contact details must agree everywhere they appear. An agent that finds a different founding year on your About page than in your Organization schema has no principled way to choose, and inconsistency of this kind is a reason to discount a source rather than merely a cosmetic flaw.
This is unglamorous and it is frequently where most of the real work is. On a several-hundred-page catalogue it is not a task anyone does by hand.
4. Format for extraction
- One H1, then a logical H2/H3 outline. This lets a system extract the relevant section rather than processing the whole page.
- Answer first, then elaborate. A section that opens with a direct statement is far easier to extract cleanly than one that builds to its point.
- Lists for sequences and feature sets. Ordered for steps, unordered for characteristics.
- Real tables for tabular data — genuine
tableelements with proper header cells, not CSS-styled divs that carry no semantics. - Define your terms. Do not assume domain knowledge that a general-purpose system will not have.
- Do not hide content behind JavaScript. Content that only exists after client-side rendering may not be seen at all.
5. Maintenance
Structured data that goes stale is worse than none, because it is confidently wrong. An agent will repeat a discontinued product or a superseded price with exactly the confidence it would repeat a correct one, and you will never see the conversation in which it happened. Price, availability and opening hours decay fastest.
A quarterly audit of cornerstone content is a reasonable baseline; anything with a price on it needs a tighter cycle than that.
A note on the syndication advice
The earlier version of this page recommended repurposing your content onto LinkedIn, Medium, GitHub, Stack Overflow, G2, Capterra, Reddit and industry forums, on the reasoning that these are "platforms with high AI crawl rates" whose content enters training datasets.
That reasoning does not hold. Publishing to a third-party platform does not place anything in a model's training corpus, and no one outside the model developer can say what does.
There is a narrower version that is true: where your customers genuinely research on a platform — a review site in your category, a community where your product is actually discussed — an accurate, non-promotional presence there means an assistant retrieving that page finds correct information about you. That is worth doing for the same reason it was worth doing before AI: it is where people are. It is not a mechanism for getting into a model, and posting into developer communities as a marketing tactic is likely to cost you more in goodwill than it returns.
What to measure
- Retrieval. Can AI systems fetch your pages at all? This is a yes or no, and it is where to start.
- Citation. How often assistants name you on queries that matter in your category, and in what tone. Sentiment moves measurably — Selleys went from 82% to 95% over three months.
- Referrals. Agent traffic and AI referrals in your own analytics. This is the only one of the three tied to revenue, and it is the slowest to move.
Structured data validation belongs alongside these as a hygiene check rather than a performance measure — it tells you your markup parses, not that anyone read it.
What structured data will not do
It makes your facts retrievable. It does not make them true, complete or competitive. A thin specification published in perfect JSON-LD is still a thin specification. An incorrect price in perfect JSON-LD is an incorrect price that agents can now find efficiently. The quality of the underlying facts is the input to this work, and it is the variable most often overlooked.
And none of it guarantees citation. Retrieval can be made possible; selection belongs to the model, and depends on the query, the category and who else is competing for the same answer.
Where to start
The free AI visibility audit at norg.ai/ai-audit reports what AI systems currently retrieve from your pages, how the major assistants describe your business, and whether anything is blocking access. No meeting, nothing installed. If it finds a blocker, that is likely the most valuable thing you will learn this quarter, and it costs nothing.
Pricing is published at norg.ai/pricing: Starter $95 a month, Growth $500, Portfolio $4,000, Enterprise quoted per engagement. Australian dollars.
Publisher
Norg Pty Ltd, ABN 44 669 712 494, ACN 669 712 494. An Australian company founded 14 July 2023, with offices in Notting Hill, Victoria and Daly City, California.