Recommendation Intelligence Research™ · Study #24

It recommends you first. By turn four, it's moved on.

We took 10 brands already measured in the Recommendation Reports series and ran a new kind of conversation: an explicit, favorable introduction of the brand, then three follow-up questions that never name it again, what else is out there, what stands out, and finally, pick just one. Across 200 four-turn conversations, the brand survived to the final answer 49.5% of the time, cohort-wide. That headline number hides almost everything interesting. Three brands survived every single run. Two never survived once. And one brand with a perfectly respectable one-in-four baseline recommend rate got swapped out for the same competitor in 18 of 20 conversations.

49.5%T4 survival rate, cohort-wide
10Brands, 20 conversations each
200Four-turn conversations run
36.8ppAbove a similarly designed industry test
Where this comes from

Four turns, and the brand's name only comes up once

Every prior study in this series that touches conversation dynamics has looked at stability between separate sessions, ask the same question a week apart, does the model still agree with itself. This one looks inside a single conversation. Turn one is an explicit, favorable introduction: the brand is named and described positively, the kind of opening line a shopper who already likes the brand might use. Turns two and three never name the brand again, an open "what else should I consider" question, then a narrowing "what stands out to you" question. Turn four forces a single final pick. A cheap case-insensitive substring check flags whether the brand comes back up unprompted at T2 or T3, and a dedicated judge call classifies the T4 answer as SURVIVED, DISPLACED, or AMBIGUOUS, with AMBIGUOUS counted conservatively as not-survived in the headline numbers.

The correlation check against each brand's real-world performance reuses data rather than re-collecting it: the baseline recommend rate for each brand comes straight from its published Recommendation Report, hundreds of real buyer-question observations already collected and scored. Nothing new was gathered for that half of the comparison, only the T1-T4 conversations themselves are new calls. Brand identities are anonymized throughout this page, labeled Brand A through Brand J, to protect the individual businesses studied; the underlying rates and relationships are unchanged.

Model: gpt-4o throughout
Design: T1 explicit intro → T2 open question → T3 narrowing → T4 forced pick
Conversations: 200, 10 brands × 20 runs
New calls: ~1,000 (4 turns × 200, plus judge calls)
T4 scoring: SURVIVED / DISPLACED / AMBIGUOUS
Baseline data: reused, published Recommendation Reports
Why T1 is deliberately favorable. The point of this design is to measure how much of a strong, best-case starting position survives ordinary follow-up conversation, not to simulate an average first message. That choice matters for reading the headline number honestly, see the limitations below.
The finding

Recognition drops, then partly recovers

T1 recognition is 100% by construction, the brand was just named. By T2, with the brand unprompted, mention rate falls to 32.5%. By T3, it climbs back to 59%, likely because "what stands out to you" is a narrower question than "what else is out there," pulling the model back toward stronger candidates. T4's forced pick lands at 49.5% survival, cohort-wide (95% CI 25.5-74.5%, n=200), well above a similarly designed industry test published elsewhere in 2026, which reported 12.7% survival on its own headline number. The gap is large, and the most likely reason is the T1 design difference disclosed above: this test opens with an explicitly favorable introduction, most naturalistic conversations don't.

Brand mention / survival rate across the conversation
Cohort-wide, n=200 conversations · dashed line is a similarly designed industry test, T1 and T4 only
Atom Foundry, this study (T1 artificially favorable)
Similarly designed industry test (naturalistic T1)

The dip-then-rebound shape is the interesting part, not just the endpoint. T2's open question loses the brand fastest, T3's narrower framing brings some of it back, and T4's forced choice settles in between.

By brand

The range runs from 100% to 0%

The cohort-wide average hides a split cohort. Three brands, labeled A, B, and C here, survived all 20 of their conversations. Two, labeled I and J, survived none, consistent with their near-zero baseline recommend rates, they rarely get considered in the first place, so an artificially favorable T1 introduction doesn't have much to protect. The rest sit in between, and one of them, Brand H, is the most interesting case in the cohort, more on that below. Brand identities are anonymized here to protect the individual businesses studied.

T4 survival rate by brand
Share of 20 conversations where the brand survived to the final pick, per brand · brands anonymized
Brand A · n=20
100%
Brand B · n=20
100%
Brand C · n=20
100%
Brand D · n=20
60%
Brand E · n=20
60%
Brand F · n=20
45%
Brand G · n=20
25%
Brand H · n=20, mid-pack baseline
5%
Displaced by the same rival, 18 of 20 times
Brand I · n=20, near-zero baseline
0%
Brand J · n=20, near-zero baseline
0%
Brands I and J sit at the floor for the expected reason. Both have baseline recommend rates near zero in the published Recommendation Reports, they are rarely considered at all in open recommendation, so a favorable T1 introduction has almost nothing to build on before the conversation moves past it. That is a Candidacy problem this study can't fix, not a new finding about either business.
A follow-up question

Does a stronger baseline mean better survival?

Correlating each brand's T4 survival rate against its published baseline recommend rate gives r = 0.68 (p = 0.048, n = 9), Brand I excluded since its near-zero baseline makes the ratio undefined. That's on the edge of significance and moves in the opposite direction from the null result this series found for possession-vs-deployment, a stronger real-world track record does seem to buy some protection against mid-conversation displacement. But the relationship is loose, not tight, and the biggest miss in the cohort is the most telling data point.

T4 survival rate vs. baseline recommend rate
9 brands, Brand I excluded · dashed line is the fitted trend · brands anonymized
Baseline recommend rate, published Recommendation Reports →
Brand H is the outlier that makes the story. Its baseline recommend rate sits comfortably mid-pack, better than half the cohort. But its T4 survival rate is just 5%, one survived conversation out of 20, with 18 displaced and 1 scored AMBIGUOUS. In conversation after conversation, the model's T2 answer to "what else should I consider" defaults to the same well-established rival in its category, and by T4 that rival wins even when Brand H was the brand the conversation started with. A decent real-world recommend rate did not protect it from a strong category-default competitor once the conversation moved past the opening line. That is a cleaner example of within-conversation displacement than the cohort average alone would suggest.
Why it matters

Stability isn't just between sessions

The lock-in study elsewhere in this series showed that once a model settles on a recommendation, it tends to repeat that same answer weeks later, a between-session kind of stability. This study measures the other axis: within a single conversation, does an early, favorable mention survive to the end. The answer is that it depends heavily on which brand, and specifically on whether there's a strong category-default competitor waiting in the wings. Brands A, B, and C held their ground through every follow-up. Brand H, despite a respectable baseline, did not, because the model has a go-to answer for its category that isn't Brand H, and three follow-up questions were enough to surface it. Getting mentioned first is not the same as staying mentioned.

What this doesn't prove

The T1 design choice caps what this shows

Stating the limits up front. T1 is an explicit, favorable introduction of the brand, a deliberate best-case starting condition, not a naturalistic first message. That is almost certainly why the 49.5% survival rate here reads so much higher than a similarly designed industry test's 12.7%, the two studies are not measuring the same starting point, and the comparison should be read as context on the shape of the effect, not a contradiction of that result. The same model, gpt-4o, runs the conversation and judges the T4 verdict, a real self-grading risk. The mitigation is a manual spot-check: 17 judge calls checked by hand against the raw T4 text, spanning the full range from 100%-survival brands to 0%-survival brands to the Brand H outlier and its one AMBIGUOUS case, and all 17 matched. That's reassuring on a sample this size, not conclusive.

The H3 correlation is n=9, genuinely underpowered, p=0.048 is a real but fragile signal, not a settled fact. One conversation was scored AMBIGUOUS rather than forced into a binary; it's counted conservatively as not-survived in every headline number, which if anything understates survival slightly. Ten brands, one category mix, consumer ecommerce, one model. Cross-platform and cross-model versions of this question are a separate, larger, and currently API-key-blocked follow-up. Brand identities are anonymized throughout this page and do not affect any of the reported rates or statistics.

Supporting evidence

Two checks that the setup was clean

17 of 17 Manual spot-check agreement

17 judge verdicts checked by hand against the raw T4 conversation text, spanning full-survival brands, zero-survival brands, and the Brand H outlier including its one AMBIGUOUS case. All 17 matched the underlying text exactly.

1 of 200 AMBIGUOUS, scored conservatively

Only one of 200 T4 answers genuinely hedged between two brands rather than picking one. It's counted as not-survived in every headline number here, the conservative choice, so the real survival rate is if anything a hair higher than reported.

Find out if your brand survives the follow-up

Free AI Commerce Score™ in 10 seconds.

Getting mentioned first is only half the picture. The fastest way to see where your brand actually stands once a real conversation keeps going is a free scan, not a guess.

Free · No signup · Results in 10 seconds
Keep reading

The rest of the research series

This extends the Stability question from between-session lock-in to within-conversation displacement, and sits next to the possession-deployment gap measured in the previous study.