Does AI recommend the same brand when you ask again?

ChatGPT Search changed the first recommendation in 34,847 of 77,946 consecutive pairs, or 44.7%. Google AI Mode changed it in 35,359 of 83,077 pairs, or 42.6%. The difference was 2.1 percentage points.

Both engines need repeated measurement. BrightEdge's cross-engine disagreement result measures a different layer; this split shows that repeat-run instability also exists inside each engine.

The first recommendation changed in 2,735 of 6,274 same-day pairs, or 43.6%; 19,936 of 46,748 one-day pairs, or 42.6%; 43,076 of 98,745 pairs two to seven days apart, or 43.6%; and 4,459 of 9,256 pairs at least eight days apart, or 48.2%.

The headline is not only a long-window effect. Rerunning a prompt within a day does not remove recommendation variation, while the longer-gap group shows a difference worth watching rather than proof that the gap caused the change.

A prompt history is one prompt's repeated answers on one engine. Among 14,990 prompt histories with at least five consecutive pairs, 6,006, or 40.1%, changed the first recommendation in more than half their pairs. Only 3,429 histories, or 22.9%, never changed it.

Instability is not confined to a few prompts. Track a brand’s share of repeated recommendations instead of assigning one fixed rank per prompt.

Among industries with at least 1,000 consecutive pairs, Sales and Marketing changed in 1,327 of 2,695 pairs, or 49.2%. Consumer Electronics changed in 340 of 1,173 pairs, or 29.0%. The observed range was 20.3 percentage points.

Use an industry benchmark before treating a brand's rate as unusual. The spread identifies where to investigate and does not show that industry caused the difference.

Among 186 brands with at least 100 consecutive pairs, Ahrefs lost the first recommendation in 102 of 134 pairs, or 76.1%. Cisco lost it in 86 of 133 pairs, or 64.7%. LinkedIn lost it in 402 of 2,696 pairs, or 14.9%, and Wise in 40 of 312 pairs, or 12.8%.

These are audit starting points, not brand-quality scores. Prompt mix differs by brand, so compare a brand with its own repeated history before comparing it with another brand.

The first recommendation switched from Dynatrace to Datadog 106 times across 25 prompts and both engines. It switched from Datadog to Dynatrace 97 times across 23 prompts, and from Grafana to Datadog 87 times across 28 prompts.

Repeated switches reveal competitive sets worth reviewing. They do not prove a universal head-to-head winner because the prompts and answer contexts differ.

The primary change rate was 43.6%. It was 43.7% across 153,587 pairs where both answers named multiple brands and 43.3% across 151,767 pairs no more than seven days apart. A second way of counting reproduced the 70,206 changes in 161,023 pairs exactly. Of 414,555 answers in the window, 199,205 did not have exactly one clear first recommendation; that count included one tied first position.

Neither single-brand answers nor longer repeat gaps determined the headline. The study measures how often the first recommendation stayed the same, not recommendation quality, factual accuracy, or causation. Different names for the same brand were merged before comparison, and labels that read as product names were left out of the public leaderboards.

A switch from first to second position counts as a changed top recommendation in this study, even if the same two brands appear in both answers. That is a different business signal from a brand disappearing from the answer entirely. Track appearance rate, first-position rate and the other brands named as separate measures.

For a worked example, suppose a brand is first in six of ten answers, second in three and absent in one. It appears in 90% of answers and is first in 60%; neither figure is captured by the final screenshot alone. These are illustrative counts, not additional study observations.

The first recommendation changed in 43.6% of consecutive pairs, and 40.1% of prompt histories with at least five consecutive pairs changed more often than they held.

Run important prompts repeatedly on each engine. Report how often a brand appears and how often it is named first across the full prompt set. Treat a single answer's rank as evidence from that one answer, not a durable benchmark. Rerun the fixed-window study next quarter before treating these rates as permanent engine characteristics.