The Answer Moved Before the Evidence Did

A single before-and-after reading around EASD will not separate a congress effect from the movement the answers show without one.

The 62nd Annual Meeting of the European Association for the Study of Diabetes opens in Milan on 28 September and runs to 2 October, according to the association’s own meeting page. Its data will then enter the sources AI engines read, and a comparison of one reading before with one after will assume the answer holds still when nothing happens. In the month before this congress, it did not.

One engine’s answer stepped sixteen points as a new clinic page entered its sources

In the six cycles to 4 September, Perplexity named the medicine in 116 of 318 answers (36%). In the six from 9 September it named it in 165 of 317 (52%). No cycle in the first half read above 39%; none in the second read below 46%.

The instrument did not change: between 22 August and 23 September, one question set, all but identical across cycles, went to the same four engines twelve times about a GLP-1 medicine in the UAE weight-management category.

The change we can see is in the engine’s sources. A Dubai clinic published a page comparing the licensed weight-loss medicines on 27 August and updated it on 3 September, by its own dating. It first appeared among Perplexity’s sources in the cycle of 4 September and stayed there, and the answers that drew on it named the medicine. The step was concentrated in questions asking which option to choose. No other engine cited the page. The timing coincides; we do not claim causation.

The four-engine total did not register it

The other three engines did not follow, and the total moved three points, from 44% to 47%. With the instrument held constant, that total ranged from 41% to 51% across the twelve cycles. The clearest movement in the series sat inside the total’s ordinary range and was visible only engine by engine.

A single engine moves further still from one cycle to the next. Gemini named the medicine in 26 of 53 answers on 14 September, in 40 of 54 two days later, and in 33 of 54 five days after that. A comparison of any two of those cycles would have reported the swing as a finding.

A medicine already at the ceiling cannot show a gain at all. The Singapore medicine was named in between 98.7% and 100% of scored answers in every one of fourteen cycles. Any EASD effect on it will show, if anywhere, in how it is positioned against the medicines named beside it.

Without a baseline, a post-congress reading will credit the congress with movement that happened without it

This month’s record shows the attribution failing before a single EASD abstract is presented: one engine moved sixteen points alongside a local clinic’s page, and another moved twenty-five points between two cycles, a swing we have not attributed to any source. A post-congress reading that cannot separate movements like these from the congress is reporting the date of the reading, not the effect of the data.

The step landed where it matters commercially: in the questions that ask which option to choose. The source that arrived with it was a clinic’s patient-facing page, which no brand team writes and no compliant brand team could direct. In this market, pages of that kind are part of what the engines draw on when they recommend.

Two things taken before the congress make its effect readable: the range each reading covers when nothing controlled changes, and a record of what each engine cited. A movement outside that range, on congress sources, is a congress effect. A movement inside it is ordinary variation.

Four functions need the baseline before they report a congress effect

Brand and Commercial teams in diabetes and obesity, before they report a post-congress gain: a share-of-answer claim without a band is a claim about noise. Medical Affairs, because whether congress data reached an answer can be checked source by source rather than inferred from a total. Competitive Intelligence, because a competitor’s gain in one engine may be a clinic’s page rather than a readout. And Market Access in the Gulf, where local clinics’ comparison pages are part of what the engines read when a patient asks what to choose.

We will report the post-congress reading, including if nothing moved

Both series run through the congress. Once it closes, we will report whether any engine moved outside the range above, and whether congress sources or local pages stand behind the movement.

How these readings are taken is set out on our methodology page. To baseline your category before its next evidence event, request a briefing.