Why did your AI visibility score go up but your citations did not?

A page score and a citation count measure different things, and the gap between them is usually the point, not a fault. A score of the kind most tools produce, including ours, reads the condition of a page: does it answer the question in the words buyers use, does it have a liftable answer near the top, does it carry sourced claims. Citations are an outcome of the engine's retrieval decisions, its memory of your brand, and what every other page competing for that answer looks like.

So there are three honest explanations to work through before you conclude the work is not paying, and they are worth separating because they lead to different actions.

The commonest is that the score moved the part of the system that was never the bottleneck. If your pages were not being retrieved at all, improving the prose inside them changes nothing until retrieval is fixed.

Next, your citations may not have stayed flat at all, and your measurement cannot tell. Week to week variation in AI answers is large, and a single monthly check has a wide enough error bar to hide a real improvement or invent one.

The third possibility is a mismatch in what the two numbers cover: a score blended across engines beside a citation count from one of them, or the reverse. A rise driven by Gemini and a citation count tracked on ChatGPT are two unrelated readings sitting in one report.

What is almost never the explanation is that the page work was wasted. The lever is real and it is smaller than the way this category sells it, which is a different complaint.

Does a page score predict how often AI engines will cite you?

A page score does not predict citation volume, and the best identified study in this area is the reason. Chen and Koudas, in work submitted on 10 October 2026, collected 158,099 answers carrying 468,037 citations over two 90 day periods in contrasting markets, compared pages within the same website rather than across websites, and replayed the same questions through the engines to observe what retrieval returned. Which pages the engine's own search retrieves, for the sub-questions the model writes for itself, explains about 32% of the variation in what gets cited. The properties of the pages explain 2.8%.

A second channel sits outside the page entirely. The same study found that ChatGPT includes brands it already knows in the searches it runs, favours sites it recognises, and names brands with no supporting page at all. In controlled tests, changing only the brand name in front of identical product facts changed the engine's rankings. The authors put it as pages being cited when searched and brands also being named when remembered.

That has a direct consequence for how to read your report. A page score is a measurement of whether a page is in a fit state to be used if it is retrieved. It is not, and cannot be, a forecast of how often it will be. Any vendor presenting a score as predicted citations is claiming something the measurement does not support, and we would rather say so on our own page than have a client find out from a paper.

One qualification belongs with the figures. That study covers commercial questions in two markets with a replication among small B2B vendors. It is one study, it is unusually well designed for this field, and nothing has replicated it independently yet.

How much of the change you are seeing is just measurement noise?

More of the change you are seeing is measurement noise than almost any report admits. Schulte, Bleeker and Kaufmann ran four AI search engines daily across four campaign verticals from 24 January to 20 March 2026, eight prompts per vertical, with a second dataset of identical prompts submitted repeatedly on the same day so that run to run randomness could be separated from drift over time.

The sets of sources cited from one day to the next overlapped by only 34% to 42%. Rank order was less stable still. Brand sets held up better at 45% to 59%, with wide variance between verticals. Their bootstrap convergence analysis found that the standard error on a per brand detection rate drops below 0.10 at seven runs per prompt and below 0.08 at eight, and that source level coverage converges more slowly than brand level coverage. Their recommendation is at least seven runs per prompt per day for brand monitoring and eight where source coverage matters, and their summary line is that answers vary across runs, prompts and time, which makes one-off observations unreliable.

Hold that against the typical reporting cadence and the problem is obvious. A monthly check of twenty prompts, run once each, produces a citation count whose month to month movement is within the instrument's own variation. You can watch that number fall while the underlying position improves, and the report will contain no way to tell.

The practical fix is not a better dashboard. Report a range from repeated runs instead of a single point, state the run count beside the figure, and treat any single period movement smaller than that range as nothing happening. Our page on why two tools report different numbers for the same brand covers how this plays out between vendors.

Can one score work across ChatGPT, Gemini, Perplexity and Google at once?

One number across four engines compresses four weakly related outcomes, and there is now a published measurement of how weakly related they are. Benjamin Tannenbaum collected 589 citation observations across ChatGPT, Microsoft Copilot, Google and Perplexity from 15 fixed commercial prompts on 6 June 2026, covering 528 unique URLs and 356 domains. Mean pairwise overlap of cited URLs for the same prompt was 0.0079 by Jaccard similarity, the median was zero, and 84.9% of engine pairs shared no cited URL at all.

Two limits belong with those numbers, and we would rather state them than quote the figure cleanly. Fifteen prompts on a single day is a small, brittle sample, and the author's company sells engine measured visibility, which is the alternative the paper argues for. Read alongside the day to day instability above, a single reading per engine cannot fully separate how much the engines disagree with each other from how much each disagrees with its own next answer. The sensible reading is that 0.0079 is an upper bound on cross engine disjointness rather than a point estimate.

The direction survives both caveats, because that figure is far below anything within engine variation alone could produce. The engines are drawing from largely different pools. Chen and Koudas found the same thing from the other side: ChatGPT only cites pages its own search retrieves, and those overlap little with Google's or Bing's results.

So a blended score is a convenience, not a prediction. If you report one number, report the engine breakdown underneath it, and do not let an improvement on one engine be presented as an improvement in AI visibility. Our page on which AI engine to optimise for first covers how to choose where to concentrate.

What does a page score actually buy you, then?

A page score buys you the one page level lever that survives controls, plus a work queue. Both are worth paying for and neither is a citation forecast.

The lever is alignment between your page's language and the question being asked. Discovered Labs, analysing roughly 2 million citation observations across 10,000 cited pages, measured prompt to content alignment at a standardised beta of +0.37, with a 95% confidence interval of +0.33 to +0.41 and a false discovery rate q below 1e-73, with one standard deviation more alignment associated with roughly 30% more citations. Predictors in that study had to survive four of nine robustness checks, and Discovered Labs sells services in this category, so both the rigour and the interest belong in any honest citation of it. The same analysis found no significant independent effect for schema markup or Core Web Vitals, and small coefficients for the format furniture this category sells hardest.

That is why our own model puts 35% of its weight on alignment, more than twice any other dimension, which our methodology page sets out with the weights. It is also why a score moving up is worth something even when citations have not moved: the page is now in a condition to be used, and the remaining gap is retrieval or familiarity.

We have published what that looks like on one of our own pages. Our teardown of our own page, from 28 to 71, shows the changes and what they were for. The honest framing of a score improvement is the one in that piece: the page got better at being quotable, and being quotable is necessary rather than sufficient.

What should you measure instead of, or alongside, a score?

Measure the three stages separately, because a blended number cannot tell you which one is stuck. The table sets out what to watch, what it answers, and what it cannot.

What to measure	Question it answers	Reporting rule
Appearance in cited sources, per engine	Is the engine retrieving us at all	Several runs per prompt, reported per engine, never blended
Named in the recommendation, per engine	Does the engine prefer us once it has read us	Track separately from citation; the two move independently
Which competitors are named instead	Who we are actually competing with in this channel	A list, not a share figure, so it survives the noise
Page condition score	Are our pages in a fit state to be used if retrieved	A work queue, not a forecast
Self-reported attribution on enquiries	Did anybody arrive because of this	One question on the enquiry form; nothing else sees it

The last row is the one most teams skip and the only one that connects the work to pipeline. Kevin Indig, comparing GA4 data across 12 sites and running 632 timing comparisons across four brands, published 19 October 2026, found total Direct traffic up 140.9% between January and September while engaged Direct moved by -1.0% on monthly counts, and concluded he could not detect hidden AI traffic in analytics without self-reported attribution. Asking new leads where they first heard of you costs one form field and is the only instrument either of us has.

If you want a second reading on a report you already have, our free visibility assessment runs your buying questions across all four engines and shows which stage the gap is actually in. Our page on how to tell whether your AEO is working covers what movement looks like on each stage.

What else do people ask about AI visibility scores and citations?

The five questions below come up when somebody is holding a report where one number went up and the number they care about did not.