What 500 Audits Reveal About LLM Citations
Across 500 sites audited by Mirok, page structure carried more weight than content volume in the likelihood of being cited by an LLM. A 600-word page broken into self-contained blocks gets picked up more often than a 3,000-word article written as continuous prose. That's the clearest finding of the study, and it contradicts the prevailing assumption that you need to "write more" to show up in generative answers.
The study covers three families of sites: B2B SaaS (54% of the sample), agencies and consultancies (28%), and software vendors and technical platforms (18%). Most sites are English- and French-language, with revenue between €500,000 and €50 million. Three engines were observed: ChatGPT (with web search enabled), Perplexity, and Gemini. Citations were counted across a panel of 40 queries per site, split between decision queries ("best tool for…", "X vs Y"), purchase queries ("pricing", "cost", "how much does"), and definition queries ("what is…").
The counting method is straightforward: for each query, we log the domains cited as sources, identify the exact page when the engine exposes it, then map that page to an editorial format. A site counts as "cited" as soon as one of its pages appears at least once across the full panel. Detailed results by site type are available in our analysis of the 500 Mirok audits on B2B SaaS sites.
Methodology: How a Page Becomes Citable
Six criteria were measured on every cited page, then compared against pages on the same site that never surfaced. The first is the presence of a self-contained answer: a paragraph or block that answers a question without depending on the rest of the page. The second is the presence of dated, quantified data (prices, percentages, volumes, timelines). The third is the density of named entities: product names, competitors, technologies, standards.
The next three criteria are technical. Freshness first: a visible update date that's consistent with the content. Crawler accessibility next: server-side rendered content, no blocks in robots.txt, no login wall on key pages. Markup last: hierarchical headings, lists, HTML tables, structured data where it makes sense.
Three caveats keep this honest. The 500-site sample isn't representative of the entire web: it skews toward companies already investing in their visibility. LLM answers vary from one day to the next, one session to the next, and sometimes one user to the next: a citation rate is a trend, not a constant. And no engine publishes its proprietary source-selection data: we're observing strong correlations, not official rules.
Format Rankings: Observed Citation Rates
| Page format | Observed citation rate | Share of sites that have one |
|---|---|---|
| Comparison and versus pages | 68% | 41% |
| Pricing pages | 61% | 79% |
| Technical documentation | 57% | 33% |
| Case studies with hard numbers | 52% | 46% |
| Glossaries and definitions | 49% | 19% |
| Alternative pages ("alternative to X") | 44% | 27% |
| Structured FAQs | 38% | 52% |
| Generic blog posts | 21% | 94% |
The immediate read: the two most-cited formats aren't the most common. Comparison pages lead the pack even though fewer than half of sites publish them. Pricing pages are everywhere but only convert into citations when they're extractable. Technical documentation, cited in 57% of cases, exists on just a third of the sites in the sample.
Conversely, the generic blog post is the most-produced format (94% of sites) and the least cited (21%). The problem isn't blogging, it's unstructured blogging: the posts that earn citations are the ones containing a comparison section, a table, an isolatable definition, or a dated figure. The declared format matters less than the internal structure.
The direct consequence for an editorial strategy: stop producing volume in the hope that a page sticks, and start producing formats with a clear intent. A site that adds three comparison pages and one readable pricing page does more for its AI visibility than a site that doubles its publishing cadence. The gaps between formats are wide enough to justify an explicit trade-off in the editorial calendar.
Comparison and Versus Pages: The Most-Cited Format
Comparison pages dominate because they answer a decision intent, the one that precedes a purchase. When a user asks ChatGPT "which tool should I choose between A and B," the engine looks for a page that lays out the criteria, compares them, and takes a position. A page that only describes its own product doesn't provide comparable material: it gets skipped in favor of third-party content that did the work of putting options side by side.
The winning structure is almost always the same: a readable HTML table, criteria named as columns, one row per solution, and a verdict per use case rather than a single winner. Pages that declare "we're the best" without explicit criteria earn citation rates close to those of generic articles. Pages that acknowledge limits ("less suited if you need X") get picked up more often, because they provide a usable trade-off.
The fix comes down to four points. First, replace screenshots of tables with real HTML tables that a crawler can read. Second, explicitly name the comparison criteria (price, plan limits, integrations, compliance, setup time). Third, display a visible update date and keep it current, because a stale comparison loses its value as a source. Fourth, write a verdict per usage profile, with a one-line sentence that summarizes the choice and can be isolated. That last point ties into the principles detailed in our guide on the 8 GEO fixes for structuring your pages.
Pricing Pages and Technical Documentation: The Sources AI Prefers
The pricing page is the factual source of record. Engines look there for plans, quantified limits, units (per user, per month, per event), and terms (commitment, overage, annual discount). A well-built pricing page is cited in 61% of cases on purchase queries, making it the second most-cited format in the study.
The most common problem found in the audits is impossible extraction: prices loaded via JavaScript, shown in a dashboard after login, or buried in an image. A crawler that can't see the price can't cite it. The fix is to render prices in the served HTML, pair them with an update date, and expose each plan's limits, not just its promises. A page that says "up to 10,000 events per month, then €0.02 per additional event" is infinitely more citable than a page that says "flexible pricing."
Technical documentation plays the same role on a different turf: it's the source of truth for prerequisites, commands, versions, and compatibility. It's cited in 57% of cases, often on queries the marketing team never considered commercial. The fix is structural: break the docs into question-and-answer blocks, one heading per problem solved, one command per block, a version and date per page. Doc pages that open with three paragraphs of context before reaching the command are routinely ignored in favor of a third-party forum or GitHub repo.
Case Studies With Numbers and Glossaries: The Underused Formats
These two formats show high citation rates for a very small footprint in the sample. Case studies with hard numbers are cited in 52% of cases even though fewer than half of sites publish them, and glossaries in 49% of cases even though they exist on only 19% of sites. This is the biggest opportunity pool identified by the 500 audits.
A citable case study rests on three elements: dated figures, an identifiable client context (industry, size, stack), and a measurable before-and-after. "We helped a client improve" produces zero citations. "Over 60 days, a 40-person SaaS vendor went from 3 to 17 citations in Perplexity across 40 purchase queries" is a sentence engines can lift as-is. The fix is therefore editorial before it's technical: enforce a dated-numbers format and a context box at the top of every case study.
Glossaries work differently. They capture definition queries, which make up a significant share of the questions people ask LLMs. A winning definition is short, self-contained, starts with the term in bold or as a heading, includes related entities and synonyms, and ends with a sentence of context. A glossary page grouping 40 definitions of 60 words each, each with its own anchor, generates more citations than a 4,000-word deep dive on the same topic. The fix is simple: create the page, one definition per block, clean markup, and link each definition to the relevant product page.
The Fix Table, Format by Format
| Format | Most common problem detected | Fix to apply | Estimated effort |
|---|---|---|---|
| Comparison and versus pages | Tables as images, no verdict | HTML table, named criteria, verdict per use case, update date | 4-6 hrs per page |
| Pricing pages | Prices loaded via JavaScript or hidden | Prices in served HTML, quantified limits, update date | 2-4 hrs |
| Technical documentation | Context before the answer, no version | Q&A blocks, command up top, version and date per page | 1-2 hrs per page |
| Case studies with numbers | No dated figures, vague context | Dated figures, client context, measurable before-and-after | 3-5 hrs per study |
| Glossaries and definitions | Format missing from the site | 30-50 definitions of 60 words, one per block, entities and synonyms | 6-10 hrs total |
| Alternative pages | Copy of the comparison page without criteria | Explicit criteria, cases where the alternative is preferable | 3-4 hrs per page |
| Structured FAQs | Questions in non-indexable accordions | Questions as headings, self-contained 2-4 sentence answers | 1-2 hrs |
| Generic blog posts | Continuous prose with no isolatable block | Insert a table, a definition, a dated figure per post | 30 min per post |
This table is the most directly reusable part of the study: it turns a visibility finding into a task list. In practice, sites that apply the first three rows (comparisons, pricing, docs) cover the majority of their market's purchase queries. The remaining rows broaden coverage toward definition and substitution queries.
One methodology note to close: these fixes shouldn't be applied blindly. Every page change should be validated before publishing, then tracked to measure the real effect on citations. That's exactly the protocol described in our article on automation with human validation: detection, audit, sourcing, validation, execution.
How to Audit Your Own Formats in 7 Days
The plan comes down to five steps, one per business day, with two days of buffer for analysis.
Day 1: inventory your pages by format. List every indexable URL and sort them into eight categories (comparison, pricing, docs, case study, glossary, alternative, FAQ, blog). You'll immediately see which formats are missing, and that's often where the fastest win is.
Day 2: build a panel of 20 purchase queries. Ten decision queries ("best tool for…", "X vs Y"), five pricing queries, five definition queries. Phrase them the way your customers would ask, not the way you'd like to be searched for.
Day 3: run those 20 queries in ChatGPT, Perplexity, and Gemini, and note who gets cited in your place. The result is often uncomfortable: you discover that three competitors and two directories occupy the space you were targeting.
Day 4: cross-reference the results with your inventory. For each query where you're not cited, identify the format that won and check whether you have that format. Missing formats explain most absences.
Day 5: prioritize by commercial intent. A missing comparison page on your core market comes before a glossary. Size the effort using the table in the previous section and kick off the first two or three projects.
Days 6 and 7 are for setting up tracking: without repeated measurement, you won't know whether your fixes are working. That's the role of continuous monitoring, which re-runs the query panel and flags movements. To move faster, the "14 days to get cited by ChatGPT" challenge walks through these steps directly in the cockpit, and the Quick Win playbook lists the ten most common technical fixes detected by the agents. If you want to industrialize the approach across multiple sites, the logic is laid out in our article on AI visibility for agencies.
FAQ
Which content format gets cited most by ChatGPT, Perplexity, and Gemini?
In the study, comparison and versus pages lead with a 68% citation rate, followed by pricing pages (61%) and technical documentation (57%). Then come case studies with numbers, glossaries, alternative pages, and FAQs. The order matters less than the underlying principle: a page structured in self-contained blocks gets cited more often than a long, unstructured page.
Do I need to write longer to get cited by AI?
No. The 500 audits show that length isn't correlated with citation rate, while the presence of isolatable blocks strongly is. A paragraph that answers a question without depending on the page's context is directly reusable by an engine. A 3,000-word article in continuous prose, with no table, no isolatable definition, and no dated figure, is rarely picked up.
Can a pricing page really get cited in an AI answer?
Yes, it's one of the most-cited sources on purchase queries, with a 61% observed citation rate. The condition is technical: prices and each plan's limits must be present in the served HTML, without relying on JavaScript or a login. Add a visible update date, because a price without a date loses its value as a source.
How long before citations start showing up?
In the audits, technical reindexing of a modified page generally takes anywhere from a few days to three weeks. Citation stabilization is slower: expect one to three months to see a clear, lasting movement across a query panel. That's why continuous monitoring is essential: it measures the real effect instead of assuming a fix worked.
How do I know if my content is cited by LLMs?
Test a panel of 20 purchase queries in ChatGPT, Perplexity, and Gemini, log the sources cited, and compare them against your own pages. Repeat at regular intervals, because answers vary from session to session. Mirok's GEO monitoring automates this tracking: agents re-run the panel, detect citation movements, and flag opportunities to act on, with human validation before execution.