How is this check scored?
Every threshold on this page comes from one source: the citation-readiness scorer Gadex publishes. Per Gadex's own build, the weights below are read directly from that scorer and a test fails the build if the documented figure stops matching the applied one. A pass earns the full 18 points, a warning earns half, a fail earns none.
| Result | How it is decided |
|---|---|
| Pass | A quarter or fewer of the sentences carrying a figure lack nearby attribution. |
| Warn | Between a quarter and 60 per cent lack it — or the page contains no figures at all. |
| Fail | More than 60 per cent of sentences with a figure have no source in or beside them. |
Why does this check exist?
A model deciding whether to repeat a claim has to weigh what happens if the claim is wrong. An unattributed number is the worst case: it is specific enough to be quoted verbatim, and there is nothing on the page to hang the attribution on.
This is why the check is weighted highest. A vague sentence that gets summarised loosely costs little. A precise figure repeated without a source is how an answer engine ends up asserting something it cannot defend — so the safer behaviour, from the model's side, is to skip it.
The check reads attribution in a three-sentence window, not just the sentence itself. Real writing often puts the source in the preceding sentence, and demanding attribution inside every sentence would flag correctly-sourced prose.
How do you fix it?
- Put the source in the same sentence as the figure, or in the one immediately before or after it. A link to the source counts; so does naming it — "according to", "per Gartner", "data from".
- Where a figure is your own measurement, say so and say how you measured it. First-party data is citable precisely because the methodology is stated.
- Where a figure adds nothing, delete it. A rounded claim you can defend beats a precise one you cannot.
How often do pages fail this check?
Per Gadex's own measurement across the 91 content pages on this site, using the checker published at /tools/ai-search-visibility-checker/: 6 pass, 57 carry a warning, and 28 fail. Per that same run, 93 per cent of Gadex pages have something to fix on this check alone.
Those figures include the pages Gadex fails. A checker whose author exempts itself from the checking is marketing rather than measurement, so the numbers are generated from the built site on every deploy and the build fails if the published figures stop matching what Gadex actually scores. When Gadex fixes pages, the numbers move on their own.
What does this check miss?
- A "figure" means a percentage, a currency amount, a multiplier, a decimal, a number of three digits or more, or a smaller number paired with growth language. That trade is deliberate: matching any digit flagged list markers, durations like "a 30 minute call", and version numbers, which buried the real problems in noise.
- The cost of that trade runs the other way too — a small bare integer doing genuine evidential work ("we audited 12 pages") is missed.
- A source three sentences later is still flagged. The window is three sentences, not a paragraph.
Where does this guidance come from?
The scoring rules are Gadex's. The reasoning about why these properties matter draws on the platforms' own published guidance:
- Google Search Central: creating helpful, reliable, people-first content — on self-evident expertise, sourcing, and who the content is for.
- Google Search Central: Article structured data — on author and date fields, and what they are used for.
- OpenAI: bots and crawlers — on how ChatGPT fetches and attributes pages it surfaces.
How we measured this
The counts above come from running the Gadex scorer over the built HTML of every content page on this site, reading each page's main content region only — nav, header and footer links are excluded, because a visitor pastes an article rather than a whole document. Legal boilerplate, the 404 page and generated tool result renderings are excluded as pages where the score would be meaningless. Per that run, the 91 pages counted are the ones a reader could plausibly ask an assistant to summarise.
The scoring thresholds on this page are read from the same module the tool runs, and a test fails the build if the weight documented here stops matching the weight the scorer applies. What none of this can tell you is whether an answer engine will cite a page: citation depends on the model, the prompt, the index and the competing sources, none of which is visible in your markup.
Common questions
Does every number on the page need a citation?
No. Prices you set, counts of your own features, and figures describing your own process are yours to state. The check is looking for claims a reader could dispute and a model would have to defend — third-party statistics, market numbers, research findings.
Why is a page with no figures at all only a warning?
Because it is not a defect, but it is not neutral either. Nothing is unsourced, so there is nothing to fix — but a page with no specifics also gives an answer engine little concrete to lift and attribute.
Does a link to our own site count as attribution?
Yes, and that is a known limitation. The check cannot tell which domain the page belongs to, so an internal link reads as an outbound one. Use your judgement on the quoted sentences rather than trusting the count.