A search fan-out trend needs the same measuring instrument

Nectiv's August 13 study reports an average of 7.61 search fan-out queries, compared with 2.17 in its October 2025 study. But the newer method extracts queries through an API using GPT-5.6 Sol, while the earlier account describes extracting search data from the ChatGPT interface. The measuring surface changed alongside the model. New study, earlier method.

The increase may reflect a real behaviour change. I would not call it a model effect until the comparison separates that possibility from changes in collection and configuration. A precise average cannot perform that separation for us.

For anyone buying AI-visibility reporting, my rule is simple: a trend line needs a record of the instrument that produced each point. Otherwise a dashboard can turn a tooling change into a story about the market.

Some continuity does not remove every difference

Nectiv says the newer study used about 4,000 prompts drawn from the earlier set, with representation across its verticals. That is useful continuity. The earlier analysis used more than 8,500 prompts, so the sampling and weighting still deserve attention alongside the interface-to-API change. New study's method.

None of those differences proves the reported increase is an artefact. They explain why its cause remains unresolved. The defensible result is that the two reported configurations produced different averages, under the methods the author describes.

That wording may sound less dramatic, but it gives the next experiment a job. Keep the prompt set fixed, run the competing collection methods over comparable configurations where possible, and record which differences remain. If a comparison cannot be made, label the discontinuity rather than drawing a seamless line through it.

Define what the counter counts

In a hypothetical visibility review, I would also ask what the average divides by. Does it include every submitted prompt, only prompts that triggered search, or only successfully captured sessions? Those denominators can produce different values even when the underlying search behaviour is identical.

As an illustration, ten prompts that trigger no search and ten prompts that each trigger four searches produce two queries per submitted prompt. Count only search-triggering prompts and the average becomes four. Neither calculation is inherently wrong; an unlabeled switch between them is.

I would require the reporting supplier to retain the model, product surface, prompt-set version, collection date, search-trigger rate and handling of failed captures. This is a proposed reporting contract, not a claim that I have audited Nectiv's raw data. The published article does not provide enough raw observations to reproduce its result independently.

The same care applies to percentages inside fan-out queries. A pattern occurring in a share of generated searches is not automatically present in that share of user prompts. One prompt can contribute several searches, which changes how much weight it receives.

For a content team, the immediate response should be to inspect relevant queries and the sources they retrieve. Expanding an entire editorial programme because a general average rose would outrun the evidence. A useful study can suggest where to look without dictating the budget.

The better dashboard would show a break when its instrument changes and begin a comparable series from that point. That is more honest than a smooth historical chart, and more useful when somebody later asks which intervention actually changed visibility.

Before explaining a jump in AI search behaviour, establish whether you changed the ruler.