Do lost AI recommendations come back?

The noise floor is how often the first-named brand stays the same from one repeat answer to the next, when nobody changed the prompt. Across 688,012 consecutive pairs it stayed the same 52.0% of the time and changed 48.0% of the time.

Google AI Overviews was the steadiest engine at 57.1%. ChatGPT kept its first-named brand 48.2% of the time, ChatGPT Search 50.1% and Google AI Mode 50.3%. Our earlier study of the #1 recommendation found a similar rate of change, about 45%, on ChatGPT Search and Google AI Mode.

The chance rate is how often two answers would name the same brand first if each answer picked its first-named brand at random from that prompt's usual mix over the period. A prompt where one brand is named first most of the time has a high chance rate even with no memory between answers.

Consecutive answers beat the chance rate by 1.23x on ChatGPT, ChatGPT Search and Google AI Overviews, and by 1.33x on Google AI Mode. On ChatGPT, for example, 48.2% of repeat answers kept the first-named brand against a chance rate of 39.0%. Most of the agreement between two repeat answers comes from the prompt's usual mix, not from the previous answer.

Writesonic reported a similar 1.18x advantage for the brand at #1 in its 631,999-prompt study. We also measured how that agreement changes as the time between answers grows.

We compared every pair of answers to the same prompt on the same engine up to 120 days apart, and divided the agreement by the chance rate. A value of 1.0 means two answers agree no more than chance.

Answers one day apart agreed 1.24x to 1.45x as often as chance. The advantage shrank every week. By 31 to 45 days apart it was 0.99x on ChatGPT, 0.98x on ChatGPT Search, 0.96x on Google AI Mode and 1.01x on Google AI Overviews, which reached 0.97x at 46 to 60 days. After that, agreement fell below chance on all four engines. That drop below chance is the real change: the prompt's mix of first-named brands itself moved over the months.

A run is a stretch of consecutive repeat answers with the same first-named brand. We followed 330,554 runs that started inside the window, after the first answer we observed for each prompt.

69.2% of new runs ended after a single answer. 16.1% lasted more than 2 answers, 3.6% more than 7 and 0.5% more than 30. The median run lasted one answer on all four engines. Taking the first spot once is common. Keeping it for more than 7 answers in a row is rare.

Comebacks depended on how long the brand had held the first spot before losing it. After holding it for one answer, the brand came back within 7 answers 51.0% of the time. After 2 to 6 answers, 80.9%. After 7 to 29 answers, 94.7%. After 30 or more answers, 97.9%.

A brand that has been named first for 7 to 29 answers in a row and then loses the spot is back within the next 7 answers about 19 times in 20.

A lasting loss is a loss after which the brand is not named first in any of the next 14 repeat answers. For brands that had held the first spot for 7 or more answers, 2.4% of 8,527 losses were lasting losses.

The rate was 2.0% on ChatGPT and Google AI Overviews and 5.8% on ChatGPT Search and Google AI Mode. The current engines are asked less often, so 14 answers cover more calendar days, which gives a real change more time to show up.

The top three are the first three brands an answer names. On consecutive repeat answers, the same three brands, in any order, came back 20.4% of the time across all four engines.

Google AI Overviews repeated its top three 25.6% of the time. ChatGPT, ChatGPT Search and Google AI Mode repeated them 15.9% to 17.8% of the time.

Month by month, Google AI Overviews kept its first-named brand 57.8% to 61.2% of the time from December 2025 to March 2026. Google AI Mode, which replaced it in our data, stayed between 49.3% and 51.8% from June to October 2026.

ChatGPT moved between 44.1% and 51.2% over the same months, and ChatGPT Search between 43.9% and 52.9%. October and November 2025 read lower on both earlier engines; brand detection in our data changed in December 2025, so treat those two months with care.

For each prompt on each engine we found its usual first-named brand: the brand named first most often over the period. On average it was named first on 53.1% of ChatGPT answers, 59.8% of Google AI Overviews answers, 54.8% of ChatGPT Search answers and 53.0% of Google AI Mode answers.

Only 10.9% of the 9,612 prompt and engine pairs had one brand named first on 90% or more of answers. In 46.0% of them, the usual first-named brand held the spot on fewer than half of answers.

The market stability score is the share of consecutive repeat answers that kept the same first-named brand, for prompts in that market. Among the 21 markets with at least 10 prompts in our panel, CI/CD and DevOps Platforms scored 75.0%, Edge Security and DDoS Platforms 74.7% and International Money Transfer Services 74.2%.

Reddit and Community Marketing Services scored 35.1%, the lowest. Knowledge Base and Wiki Software scored 47.8%. Markets with one clear default tool held the first spot best.

The longest unbroken runs covered the whole observed period on ChatGPT and Google AI Overviews. PlayStation was named first in all 131 Google AI Overviews answers to "What is the best gaming console?" from October 2025 to March 2026. Miro and Boudin Bakery held 131-answer runs on ChatGPT.

On the current engines, Patagonia was named first in 102 straight ChatGPT Search answers about sustainable clothing, and Cloudflare in 94 ChatGPT Search and 93 Google AI Mode answers about DDoS protection. Runs this long are the exception: 0.5% of new runs lasted more than 30 answers.

Brand retention is how often a brand named first in one answer was named first again in the next repeat answer. Among the 41 brands named first in at least 1,000 answers, NordVPN kept the spot 91.3% of the time, Wise 86.2%, LaunchDarkly 85.7% and Docusign 84.8%.

IBM kept it 43.2% of the time and Microsoft 47.3%, below the 52.0% noise floor. Broad brands that are named first across many different prompts held the spot less than specialists that own one question.

We kept one answer per prompt, engine and day, and took the brand named first in it. Every engine comparison uses the same 2,403 organic prompts, each answered on at least 30 days on all four engines: ChatGPT and Google AI Overviews from October 19, 2025 to March 31, 2026, and ChatGPT Search and Google AI Mode from May 29 to October 6, 2026. The two engine pairs are never mixed in one run or one pair. April 2026 is excluded because brand detection failed for most answers that month.

The first-named brand is not always the recommended brand: an answer can name a material or a platform before its pick. Our earlier study of the #1 recommendation on ChatGPT Search and Google AI Mode found a similar change rate. The current engines were asked every one to four days, the earlier ones mostly daily, so 7 answers cover more calendar days on the current engines. When two brands tied for first position, we broke the tie the same way every day. We computed the headline two ways, from an exported daily series and from a separate in-database rebuild, and both give 60.0%.

Do not react to one lost first mention. On every engine the first-named brand changes on about half of repeat answers, and 60% of losses reverse within 7 answers. For a brand that held the spot for 7 or more answers, a loss reverses 95% of the time.

Track the share of answers that name you first over consistent windows. In this study, established brands rarely stayed out of first place for 14 repeat answers—a pattern worth investigating, rather than reacting to one answer.