IndexHaloThe observable citation pipeline
A practical model has five stages: discovery, access, retrieval, evaluation, and presentation. A failure at any stage can remove a page from consideration. A brilliant report hidden behind a bot challenge has an access problem. An accessible 4,000-word article with no descriptive sections has a retrieval problem. A precise claim with no source or date has an evaluation problem.
1. Discovery: can systems find the URL?
Discovery begins with ordinary web architecture: crawlable links, accurate XML sitemaps, stable canonical URLs, sensible status codes, and a robots policy that matches the publisher’s intent. Google recommends absolute canonical URLs in sitemaps, while Bing accepts XML, RSS, Atom, and text sitemap formats. A sitemap is a discovery signal, not a guarantee of indexing.
For AI-specific access, publisher controls still matter. OpenAI documents separate user agents for search discovery, user-requested page access, and model training. Teams should decide their policy deliberately and test the production response rather than assuming a robots file behaves as intended.
2. Retrieval: can the relevant evidence be isolated?
Answer systems often need a passage, table, definition, or comparison rather than an entire page. Descriptive headings, concise opening summaries, explicit entity names, units, dates, and self-contained sentences make relevant evidence easier to retrieve. This is not an argument for robotic prose. It is an argument for reducing ambiguity.
Consider “growth increased significantly.” It lacks an entity, baseline, period, and source. “European mid-market adoption rose from 31% in 2024 to 43% in 2025, according to the linked survey of 612 firms” gives a retrieval system and a human reviewer far more to work with.
3. Evaluation: is the source defensible?
Strong evidence pages show who produced the information, how it was produced, when it applies, and where the underlying source can be inspected. Primary evidence generally reduces the number of inferential steps. If a page cites a blog that cites a press release that summarises a dataset, linking the dataset directly creates a clearer provenance path.
Consistency also matters. Visible authorship, publication dates, canonical metadata, and Article structured data should agree. Structured data is not a substitute for visible information; it should describe what the reader can actually see.
4. Comparison: why another source may be selected
Being relevant does not mean being the best available evidence. Another source may be clearer, more current, more primary, easier to access, or more directly aligned with the query. Live citation testing is most useful when it records competing domains and the exact passages those pages offer.
Create a source-gap table with the query, cited competitors, evidence type, publication date, answer format, target-page gap, and recommended response. The goal is not to copy a competitor. It is to understand the evidence standard the result set currently rewards.
5. Presentation: citations vary by answer and provider
Citations can change when the wording, location, model, provider, or date changes. A cited URL may be a canonical article, a PDF, a syndication copy, or an unexpected subpage. Save returned URLs exactly and resolve redirects so the team knows which asset actually received visibility.
Seven changes that improve citation readiness
- Make every valuable page return a clean 200 response to intended crawlers.
- Place a concise answer or finding near the top of each major section.
- Give statistics an entity, unit, period, sample, and adjacent source.
- Link primary evidence directly instead of relying on citation chains.
- Align visible authorship and dates with metadata and structured data.
- Use internal links with descriptive anchor text from relevant hub pages.
- Repeat the same query set over time and compare observations, not anecdotes.
Why a good page still loses the comparison
The most common question we get is some version of: our page is genuinely the best resource on this topic, so why is a thinner competitor being cited instead? Usually one of five things is happening, and none of them are about quality in the way the author means it.
The competitor answers the question asked. Retrieval matches passages to a query, not documents to a topic. A 4,000-word definitive guide that never states the specific answer in a liftable form will lose to a 600-word page that does. Comprehensiveness is not the same as responsiveness.
The competitor is easier to attribute. Clear authorship, a visible date, a named organisation and a linked methodology all reduce the risk of citing a source. Engines are risk-averse about attribution because a bad citation is more damaging to them than a missed one.
The claim is not adjacent to its evidence. A statistic in paragraph three and its source in a footnote at the bottom are, from a passage's perspective, unrelated. Keep the number and its provenance in the same retrievable unit.
The page renders slowly or partially. Content that requires JavaScript execution to appear may be present for a user and absent for a fetcher working under a timeout.
The topic has an incumbent. Some queries resolve to a canonical source — a standards body, a regulator, a widely-recognised reference. Displacing one requires evidence they do not have, not a better-structured version of the same information.
How to test this yourself
You do not need a platform to start. Build a set of twenty questions a prospective customer might actually type, phrased as questions rather than keywords. If you have not yet checked whether the engines can reach your pages at all, start with the free crawler access checker — a low citation rate on a blocked site tells you nothing about your content. Ask each one in ChatGPT, Perplexity and Google AI Overviews. Record, for each: whether you were cited, which URL was returned, which competitors appeared, and the date.
Repeat the same set a fortnight later without changing the wording. The comparison between two dated runs is the entire value — a single run tells you almost nothing, because these systems are non-deterministic and the same question can produce different sources on consecutive attempts.
Two disciplines make this worth doing. Keep the query set fixed, because changing the questions between runs means you are measuring the questions rather than your visibility. And archive the full response text, not just a yes or no — when a citation disappears, the archived answer usually shows which competitor replaced you and what they said that you did not.
What a citation test proves
A preserved provider response can prove that a URL was returned as a citation for a specific query at a specific time. It cannot prove that the provider trained on the page, permanently ranks it, or will cite it for every user. Reporting that boundary explicitly makes the result more credible and more useful.
This matters commercially as well as intellectually. A report claiming “we increased your ChatGPT visibility by 40%” without defining the query set, the sampling method and the confidence interval is not measurement — it is a number chosen to look like one. Ask any vendor, including us, to show the raw archived answers behind an aggregate. The tool selection guide turns that into a trial protocol you can run against any product. If they cannot, the aggregate is decoration.
Primary references
Common questions
Why does the same question return different sources each time?+
These systems are non-deterministic. Retrieval involves sampling, and the candidate set can differ between runs even for identical queries. This is why a single observation proves possibility rather than likelihood, and why presence rate across repeated runs is the only honest unit of measurement.
Does being cited in an AI answer send traffic?+
Sometimes, and less than a comparable search ranking. Some surfaces pass a referrer and appear in analytics; others do not. A user who reads an answer and later searches your brand is invisible to attribution. Expect the commercial value to show up partly as brand demand rather than as clicks.
Do I need to be in the training data to be cited?+
No, and conflating the two causes most of the confusion in this area. Live citations come from retrieval at query time — the engine fetches or looks up current sources. That is why declining training crawlers while allowing search crawlers is coherent rather than contradictory.
Does schema markup make me more likely to be cited?+
Structured data helps a machine parse what a page is about, which reduces the chance of misinterpretation, and it is well worth implementing. It is not a citation lever on its own. A page with perfect schema and unsourced claims still loses to a page with plain markup and linked evidence.
How many sources does a typical answer cite?+
Usually a small number, often three to five, drawn from a larger retrieved candidate set. This is the strategically important part: the gap between being retrieved and being cited is where most of the competition happens, and it is decided largely by attributability.