Your audit stops at turn one

A turn-one audit is an AI visibility measurement that sends one prompt, reads one answer, and stops, which is how most GEO tracking works today and not how a meaningful share of real AI search sessions actually end. The assumption buried in that design is that the opening answer is the answer. It often is not, and the evidence sits in three research directions that are rarely read together.

#A fifth of real search conversations continue

Search Arena, the open benchmark of live search-augmented model traffic, published the shape of real sessions: of 24,069 user-model conversations, 22.4% are multi-turn, with 3,288 at two turns, 966 at three, and 460 at four. The paper describes them as "typically clarifications or follow-up queries" across 13 search-augmented model variants.

That number cuts both ways, and it is worth being honest about the smaller half of it. Most sessions really do end after one exchange, so turn-one measurement is not measuring nothing. But the continuing fifth is not a random fifth. A clarification happens when the first answer did not settle the question, which describes most of the comparison and shortlist prompts that commercial GEO programs care about. The same dataset shows recommendation prompts carrying more citations per response on average than factual lookups, 6.9 against 5.7, so the prompts most likely to continue are also the ones with the most slots to lose.

#The second turn is a different measurement, not more of the same

The clearest statement of the problem is in a September 2026 survey of GEO visibility measurement. Its section on prompt design, titled "the need is not always specified in the first turn," puts it plainly: "a source can be absent from the first answer and cited after a follow-up", and therefore "a first-answer citation rate, a rate per turn, and the probability of at least one citation during a conversation therefore have different units and denominators."

Three metrics, three denominators, one dashboard column. That is the whole problem. A brand's first-answer presence rate, its rate across all turns, and its odds of appearing somewhere in a conversation can sit far apart, and nothing in a turn-one report tells you which one you are looking at. The same survey demonstrates how much definitional choices move published figures elsewhere: reweighting the same three subcorpora, without changing any rate inside them, moved a generative-answer activation rate from 39.7% to 70.5%. Denominators are not a footnote in this field. They are most of the variance.

The survey also names the missing piece in evaluation design. To score a whole interaction, a prompt set has to be extended with "a user policy" that produces the next message from the history. Your prompt list has no such policy. It has openings.

#Models get less reliable as the conversation goes on

The reason a second turn is risky rather than merely different comes from the ICLR 2026 paper on sharded conversations. Taking fully specified instructions and revealing them one piece per turn, across more than 200,000 simulated conversations and 15 models from eight providers, produced an average 39% performance drop against the single-turn version of the same task. The decomposition matters more than the headline. Aptitude fell by an average of 16%, while unreliability rose 112%. The models did not get less capable. They got less predictable, largely by committing early to an assumption and failing to recover.

Two caveats before anyone puts that figure on a slide. The six tasks were code, SQL, API calls, math, table captioning and multi-document summarization, none of which is brand recommendation. The conversations were also simulated rather than observed. What transfers is the mechanism, not the magnitude: gradual disclosure of a need produces early commitment, and early commitment is exactly how a shortlist gets fixed before your evidence arrives.

A separate March 2026 study pushes on the other side of the same behavior by presenting candidate answers sequentially to 17 models across three clinical benchmarks. End-to-end accuracy and correct abstention fell by up to 30% on average, with one model losing 65%. The detail worth stealing is what the authors call blind switching: models moved from abstaining to accepting a suggestion at near-identical rates of roughly 50% whether the suggestion was right or wrong. The domain is clinical, so treat the numbers as domain-bound. The direction is that a confident follow-up naming a competitor is likely to be accepted on its confidence rather than its correctness.

#Your own quality checks will not see the drift

There is a measurement trap inside the measurement trap. A 2026 study monitoring conversational consistency from token statistics across 4,574 turns and 34 conditions found that its structural signal aligned with structural consistency in 85% of conditions but with semantic quality in only 44%. Structural coherence and answer quality are only loosely coupled. A conversation can stay fluent, on-topic and pleasant to read while the thread that determines who gets recommended has already moved.

That is why spot-checking a few follow-ups by hand rarely catches anything. The answers look fine. Nobody reads the fourth answer next to the first one and counts which brands survived.

#What to change this quarter

Three things, in order of cost.

Add a second turn to a slice of your prompt set, not all of it. Pick the twenty prompts that actually drive pipeline, write one realistic follow-up for each (a constraint, a price objection, a competitor named by the user), and record brand presence at both turns. You now have a first-answer rate and a per-turn rate, and the gap between them is the finding.

Report the three rates separately and label the denominator in the column header. First-answer presence, presence per turn, and presence anywhere in the conversation are different products. Averaging them produces a number that describes no situation.

Test the clarification turn specifically, because it is the one the benchmark says is most common. Google documents that AI Mode answers follow-up questions and that it divides a question into subtopics and searches for each one simultaneously. That fan-out is documented against the question being asked, so the safer assumption is that a follow-up is its own retrieval problem rather than a formatting pass over the first answer. Anything that failed to be retrievable on turn one gets another chance on turn two, and anything that won on turn one has to win again.

The uncomfortable version of all this is that the industry standardized on the cheapest possible unit of observation, one prompt and one answer, at the moment the interface stopped working that way. The fix is not more runs of the same opening prompt. It is a second message.