Does AI keep the same recommendation when its sources change?

The top recommendation stayed the same in 12,243 of 25,883 back-to-back answer pairs where no cited page repeated, or 47.3013%. It changed in the other 13,640 pairs, or 52.6987%. A back-to-back pair contains two consecutive answers from the same engine to the same organic prompt.

A completely different set of cited pages did not always change the top brand. Track the recommendation outcome and the cited pages as separate fields. SparkToro found less than a one-in-100 chance that repeated ChatGPT or Google AI runs would produce the same complete brand list twice. This study narrows that question to one first recommendation and the pairs where every cited page changed.

No cleaned cited-page address repeated in 25,883 of 122,882 back-to-back answer pairs, or 21.0633%. These complete-turnover pairs covered 10,077 organic prompts.

Complete page turnover occurred often enough to audit as a recurring condition. The rate describes this observed prompt corpus and should not be treated as a probability for every future answer.

The same top brand remained first in 12,243 of 25,883 pairs with no shared page, or 47.3013%; 14,597 of 27,424 pairs with one shared page, or 53.2271%; and 43,707 of 69,575 pairs with two or more shared pages, or 62.8200%. The last group was 15.5187 percentage points above the no-shared-page group.

Page continuity was associated with higher recommendation stability. It does not show that a repeated page caused the engine to keep the brand first. Ahrefs similarly found that Google AI Mode and Google AI Overviews could reach similar conclusions while citing different pages, but its comparison was across engines rather than across back-to-back runs.

When no cited page repeated, ChatGPT Search kept the same top brand in 9,515 of 18,970 pairs, or 50.1581%. Google AI Mode did so in 2,728 of 6,913 pairs, or 39.4619%. The difference was 10.6962 percentage points.

A combined rate hides a material engine difference. Run the same source-change audit for each engine instead of applying one aggregate benchmark to both.

With no repeated cited page, the same top brand remained first in 2,879 of 6,400 one-day pairs, or 44.9844%; 8,705 of 17,994 pairs two to seven days apart, or 48.3772%; and 659 of 1,489 pairs eight or more days apart, or 44.2579%.

A longer time between answers was not associated with a steady rise or fall in this observed cut. Do not turn one interval into a general decay rule. The intervals also contain different prompt mixes.

Among displayed industries with at least 200 complete-turnover pairs, Consumer Electronics kept the same top brand in 134 of 217 pairs, or 61.7512%. Information Technology reached 57.8440%, Software 53.6052%, Financial Services 45.0063%, Commerce and Shopping 43.4343%, Internet Services 42.4171%, Food and Beverage 41.8824%, and Professional Services 39.9638%.

Compare a brand with its category baseline before treating its result as unusual. These rates describe the observed prompt mix and do not show that industry caused recommendation stability.

LinkedIn remained first in 290 of 337 complete-turnover pairs across 66 prompts, or 86.0534%. Linear reached 72.2772%, Bright Data 70.5882%, HubSpot 68.6957%, Datadog 67.2673%, Stripe 64.1975%, Microsoft 54.0816%, and Semrush 37.5000%.

Named-brand rates show where to investigate repeated prompt histories. They reflect each brand's prompt mix and do not form a universal brand-quality ranking.

The main result, using cleaned page addresses, was 12,243 of 25,883, or 47.3013%. The original, uncleaned page addresses produced 12,508 of 26,377, or 47.4201%. Requiring at least two cited pages in each answer produced 9,610 of 21,283, or 45.1534%. Retaining only the first complete-turnover pair per prompt and engine produced 5,813 of 12,630, or 46.0253%.

We excluded ambiguous first recommendations, answers without a cited page, and prompt, engine, and timestamp combinations that contained more than one answer. Page addresses were counted once after fragments, trailing slashes, and eight standard tracking parameters were removed. The study does not measure factual support, recommendation quality, user demand, or causation.

The top brand stayed the same in 47.3013% of complete-turnover pairs. Stability rose to 62.8200% with at least two repeated cited pages, and the no-shared-page rate differed by 10.6962 percentage points between the two engines.

Record the first recommendation and the cleaned set of cited pages for every repeated priority prompt. Review each engine separately. Inspect the answer before changing content because this study does not show that a source change caused the brand outcome. Rerun the fixed-window method next quarter before calling a difference movement.