Which website features predict an AI citation? What 209 vendors and 40 Perplexity prompts showed

Inside a single AI answer, the vendors whose own website gets cited have an FAQ section, a named author, deeper articles, comparison tables and visible dates more often than the vendors the same answer names without citing. Across all vendors pooled together, none of that shows up: a site scoring 90 on my readiness check was cited at the same rate as a site scoring 35, because whether an answer cites vendor sites at all is decided by the prompt, and which vendor it cites is decided mostly by brand size. That is what I found when I ran 40 buyer prompts through Perplexity on 16 October 2026, logged every vendor each answer named, scored 183 of their websites with the same tool I use for clients, and tested which checks predicted a citation. The method and the numbers are below, with the parts I do not trust marked.

By Ronan Liedmeier, fractional head of growth. Last updated 17 October 2026.

What did the test measure?

The test measured whether a vendor's own domain appeared as a source in a Perplexity answer that named that vendor, and whether a 22-check website score predicted it. The unit is a vendor mention: one company named in one answer.

An AI readiness score is a 0 to 100 estimate of how likely an engine is to retrieve, use and cite a site, built from checks a crawler can run: can the pages be read without JavaScript, do headings match buyer questions, are facts sourced and dated, is the company identifiable as an entity. Mine has 22 checks in four groups: Access, Answerability, Evidence and Entity. The free version is at growthmarketing.solutions/ai-search-readiness-check.

Setup:

A vendor counted as cited only if a source URL was on its own domain. A review site, a listicle or a marketplace page about the vendor did not count. Zoho was marked not cited in a CRM answer whose source was bigin.com, Zoho's own Bigin product site, which is the one judgment call I would make the other way with hindsight.

Does the overall score predict a citation?

No. Pooled across all 183 vendors the correlation between score and citation was -0.03 (p 0.68), and the citation rate was flat across score bands.

Two examples make the point. A US outbound agency scored 94 out of 100 and was named in an answer that cited other agencies' sites but not its own. A German bed manufacturer scored 37, with no readable article content at all, and was cited, because the answer about German mattress brands drew on its site directly.

A logistic regression of citation on the four score groups (McFadden pseudo R-squared 0.005) found nothing. Adding the category of the prompt changed that: local-service vendors were cited far more often (coefficient +1.43, p 0.003) and nothing else moved.

So the first result is about the answer. Whether Perplexity cites vendor websites at all depends on the kind of question.

Why does the prompt category decide so much?

Because the engine's answer format for a category fixes what kind of source it shows before any page is evaluated. Local prompts cite the businesses; big software categories cite the review sites that compare them.

Two of the ten local prompts returned map-style answers where every source was a business's own site. A SaaS prompt about project-management tools named ten vendors and cited three of their sites; the other sources were competing tools' blogs, Forbes, TechRepublic and a project-management blog. A prompt for ergonomic office chairs named no brand at all and cited only a comparison site.

The practical consequence: if you sell into a category where the engine's habit is to cite G2, Capterra or a trade publication, a perfect site will not change the habit. Your page has to answer a sub-question those sites do not, which is a different job from matching their format.

Which checks predicted a citation within the same answer?

Comparing vendors named in the same answer removes the category effect, and that is where the page starts to matter. Centring both score and citation on the prompt mean, over 188 mentions in the 25 prompts that had a mix of cited and uncited vendors, the correlation between total score and citation was 0.21 (p 0.004, bootstrap 95% interval 0.08 to 0.33). Cited vendors scored 4.1 points higher than the uncited vendors in the same answer.

By check, ordered by within-answer correlation:

The five checks that predict are all about the article: a question-and-answer block, a person's name on the page, enough text to be worth lifting, a table and a date. They are the things a human editor would also look for before quoting a source.

Schema did nothing. Organization structured data sat at 0.02 and sameAs at 0.07. Ahrefs reached the same conclusion with a controlled test of 1,885 pages that added JSON-LD and saw no citation uplift on AI Overviews, AI Mode or ChatGPT (Ahrefs, May 2026).

Answer-first paragraphs did nothing either, which surprised me, because Bing's guidelines say facts should be stated directly (Bing Webmaster Guidelines). My check measures whether the first sentence under a heading answers the heading, and it may be too crude: a page can be answer-first in structure and vague in content.

Within the four categories, the within-answer correlation held for SaaS (0.26, n 57), local (0.27, n 38) and ecommerce (0.25, n 31) and was near zero for agencies (0.10, n 62). Agency answers were the ones most dominated by directories and listicles, so the page had the least room to matter.

What could not be measured?

Twenty-six vendors could not be scored, and the two Access checks that passed for every vendor could not be tested at all. Both gaps bias the result.

Twenty of the 26 unscorable sites returned HTTP 429 or 403 to my checker on a slow sequential retry, and 19 of those were ecommerce stores, mostly on Shopify. Eight of the 26 were cited. A site that blocks a server-side fetch is not necessarily blocking AI crawlers, since Shopify's protection decides by IP reputation, so I excluded them rather than scoring them as failures. It does mean the ecommerce results above rest on 31 vendors rather than 50.

Every scorable vendor passed the HTTPS check and the robots.txt check, so neither could discriminate. Vercel's crawler analysis says the major AI crawlers do not run JavaScript (Vercel), and my own site failed that test a week earlier, but in this sample too few vendors failed it for the check to show a signal.

The larger confound is fame. The checker reads a vendor's own site. It cannot see Reddit threads, review-site listings or the trade press that mention the vendor, and for well-known brands that is where most citations originate. Pooled, the vendors with the most content and the most linked profiles were cited slightly less often, which is a brand-size effect: the big names are cited through third parties rather than their own pages. Within answers the same checks turn slightly positive.

What I'd do

What this changed in my own work, in order.

1. Treat the overall score as a readiness floor the site has to clear, and nothing more.
2. Demote the two checks with no supporting evidence, bot-versus-browser comparison and noindex, to information only in the next version of the checker.
3. Weight the five predictive checks up: FAQ, named author, depth, tables, dates. Re-tested with leave-one-prompt-out, the within-answer correlation moves from 0.21 to 0.22, a small gain, so the weights stay provisional until a second run on a fresh prompt set with ChatGPT and Google AI Overviews added.
4. For any client, run the prompt panel before the site audit. The panel shows whether the category's answers cite vendor sites at all. If they do not, the first deliverable is a page that answers a sub-question the review sites skip.
5. On every article: an FAQ block, a byline with an about page, a table, a visible date, 1,500 words or more. Those five are the only on-page features in this test with a measurable relationship to being cited.

If you sell B2B and want to be the source the engines quote when a buyer asks about your category, run the free readiness check on your site first. The AI Visibility Snapshot is the paid version: ten prompts in your category, your site scored on the 22 checks above, and a one-page readout in three working days.

Re-check by 17 January 2027: the second run on a fresh prompt set with ChatGPT and Google AI Overviews, whether the five predictive checks hold, whether the answer-first check needs a stricter definition, and whether Shopify stores can be scored from a residential fetch.