Survelium Docs
Docs menu

Optimization

Comparison with cited pages

For the prompts where AI leaves your brand out, your page next to the pages it cites instead, and what yours is missing.

For the prompts where your brand is missing from most AI answers, Survelium reads the pages AI cites instead and your closest page on the same subject, and lists what the cited pages have and yours does not. Each prompt becomes one row in Recommendations, labelled "Comparison with cited pages". Nothing in it uses AI: the page signals are rules, as in Crawl Checker.

Which prompts

A prompt is compared when, over the last 30 days, the brand is missing from at least half of its completed answers, all platforms together, and the prompt has at least 20 answers. Below that there is too little to go on.

The summary at the top of Recommendations counts them as Prompts where you are missing. For prompts where you appear rarely but do appear, see visibility per prompt.

The pages compared

The cited pages — the three pages AI cites most for that prompt that are not yours, whatever they are: a competitor, a guide, a publication, a marketplace. Each is marked Competitor (a confirmed direct competitor; they come first on a tie) or Other source, with how many answers cite it and on which platforms. Left out: domains you ignored, files that are not pages (PDF, images), and the same address with tracking parameters (utm_* and the like), which counts as one page.

Your page, picked in this order:

  1. The page you chose for the prompt — it always wins.
  2. A page of yours whose title and address match the prompt's words, from the pages AI cites, your watched pages, the pages AI crawlers read and your sitemap. Shown as "Suggested — confirm it or pick another page".
  3. Your page AI cites most for the prompt, often the homepage, used only when no specific page matches.
  4. Nothing — the row says "None of your pages covers this subject", with the cited pages as a model of what to write.

Utility pages (contact, cart, account, terms, privacy, tag and author archives, search) are never suggested. When AI cites the English version of a page for a prompt in the brand's language, the page in the brand's language is used, if we know its address.

What is compared

Signal On a page when
Structured data (schema.org) the page has JSON-LD
Questions and answers (FAQ) FAQPage markup, or at least three headings that are questions
Prices Offer / PriceSpecification markup, or an amount with a currency in the text
Reviews or ratings Review / AggregateRating markup, or a number of reviews in the text
A table a table with at least three rows
Lists at least two lists
Date last updated dateModified, article:modified_time or a <time> element
Author an author in the markup or the page
Length (words) —
Sections (H2 headings) —

Text in menus, headers and footers is ignored. Prices are not asked of articles (blog posts, news), only of pages that sell or present a service.

A signal is missing when your page lacks it and at least two of the three cited pages have it. For length and sections, when your page has less than half of the cited pages' median — and only if that median is worth it (300 words, 3 sections). One cited page doing something is a coincidence; two out of three is what AI gets on that subject.

The row

Open it to see:

  • how many answers the prompt had and how often you were missing;
  • your page and why it was picked, and "the same page is the closest one for N other prompts" — improving it helps them all;
  • what is missing, each with how to fix it, ticked off one by one once fixed;
  • the table: every signal, your page next to the cited pages, with links to each. A ? means the page could not be read for that signal.

Compare with another page… — choose the right page yourself (only your brand's domain). It is kept for the prompt and the comparison runs again straight away. Confirm keeps a suggested page as your choice.

Recheck my page — after changing the page, read it again now instead of waiting for next month. It counts towards the 50 manual checks a day of Crawl Checker; the cited pages are reused.

How it closes

  • Everything missing is now on your page — resolved, with a green marker on the visibility chart and the visibility before and after.
  • Your brand now appears in more than half of the answers — resolved, with "You now show up in X% of the answers", whether or not you changed the page.
  • The target moved — AI cites other pages and "two out of three" no longer holds — the row simply goes away; it is not counted as resolved.
  • Partly fixed — the row stays open, with what you fixed ticked off.
  • Dismiss a prompt you do not want to work on; the comparison will not bring it back.

How often

Once a month, for every brand. Each cited page gets a single request, and only if its robots.txt allows it; results are reused for 7 days between brands. Your page is reread when you ask for it, and every week if it is a watched page.

How we read other sites

We read other people's pages as SurveliumBot, under our own name and with a link to survelium.com/bot, and we respect their robots.txt. A cited page that blocks us, fails or is not a page is skipped for the next most cited one, down to fifth place. If fewer than two cited pages can be read, the prompt is not compared.

Where it appears

  • Recommendations — the rows, the "Comparison with cited pages" filter and the summary count.
  • Where you're missing and Missing sources — See the comparison on a prompt that has one.

View as Markdown · llms.txt