Your AI Visibility Report Is Measuring Noise

What we tested. On 29 December 2025 we sent three gift questions ("What is the best xmas gift for husband", "What is the best xmas gift for wife", "What is the best xmas gift in 2025") to three models through their APIs and repeated each question in runs of 10 answers: Gemini 3 Flash Preview (gemini-3-flash-preview), GPT-5.2 (gpt-5.2) and Grok-4-1-fast-reasoning (grok-4-1-fast-reasoning). We collected 239 answers between 15:16 and 23:34 UTC and counted mentions of 95 known consumer brands in each one.

What differed. On the husband question, Gemini averaged 14.0 brand mentions per answer in one run and 3.7 in a run that ended 65 minutes later, then 13.1 and 13.3 in two later runs. GPT-5.2 named none of the 95 brands in 24 of its 40 answers to that question (60%). Grok's count per answer varied least (coefficient of variation 23%, against 52% for Gemini). Four brands appeared in the husband-question answers of all three models: Apple, Sony, Bose and Patagonia.

Why a marketing team should repeat a measurement. The same question returned different brand counts within one afternoon, and the three models named different brands. One answer shows one point in that spread. Asking the same questions repeatedly on the models your buyers use, and keeping every answer, shows the spread and the brands that recur.

Scope. The results describe three model versions on one day in December 2025, under three English questions. Newer versions need their own measurement.

Correction note, 4 November 2026. We rechecked every figure in this article against the eight raw run files. The earlier version said GPT-5.2 returned no brands in 70% of its answers; the run files give 24 of 40 husband-question answers (60%). It said Gemini changed its recommendations four times in two hours; the fall from 14.0 to 3.7 came between two runs 65 minutes apart, and the four runs span eight hours. It described the consistency score as how often a model gives the same answer; the score is 100 minus the coefficient of variation of brand counts, a measure of spread. It gave Gemini's husband-question average as 10.27 brands and GPT-5.2's as 0.63; those were values from the first 30 answers, and the 40-answer values are 11.03 and 0.70. It listed Perplexity among the models; Perplexity was not tested. Statements about why the models behave as they do have been removed, because the study recorded behaviour and no reasons. The December 2025 results themselves are unchanged. The raw run files are public; see Data and method.

---

The Problem Nobody Talks About

You invested in AI visibility monitoring. You ran a report. ChatGPT mentioned your brand. Done?

In this test, one answer was one draw from a wide spread. Here is what the 239 answers showed.

Gemini's brand count changed between runs within the hour. One run averaged 14.0 brand mentions per answer. A run that ended 65 minutes later averaged 3.7, a 74% drop. Two later runs averaged 13.1 and 13.3. The prompt was identical in every run.

GPT-5.2 named none of the 95 listed brands in 24 of its 40 husband-question answers (60%). Its answers described gift categories such as headphones, watches and experiences. If your tracking covers GPT-5.2 alone, it covers a model that returned few named brands in this test.

Four brands appeared in the husband-question answers of all three models. Apple, Sony, Bose and Patagonia. Across the three models, 61 different brand names appeared in those answers.

A report built on one answer from one model carries this spread without showing it.

---

What Our Research Actually Measured

We collected 239 answers to three questions (gift for a husband, gift for a wife, best gift in 2025) from three models (Gemini 3 Flash Preview, GPT-5.2, Grok-4-1-fast-reasoning). Each question ran in rounds of 10 answers per model: four rounds for the husband question (40 answers per model), three for the wife question (29 Gemini answers, 30 GPT-5.2, 30 Grok) and one for the 2025 question (10 per model). We counted mentions of 95 known consumer brands in every answer, then compared counts within a model and brand sets between models.

The Findings

Grok's brand count varied least on the husband question. Its coefficient of variation was 23%, against 52% for Gemini, which gives consistency scores of 77 and 48 (the score is 100 minus the coefficient of variation). A steady count can come with a concentrated set: Amazon appeared in 38 of Grok's 40 answers to that question (95%) and Apple in 39. On the wife question Gemini varied less than Grok (scores 65 and 55), and on the 2025 question as well (78 and 54). Which model varied least depended on the question.

Gemini's brand count varied most on the husband question. Its 40 answers ranged from 0 to 20 brand mentions. One run averaged 3.7 mentions per answer against 13.1 to 14.0 in the other three, and 4 of that run's 10 answers named no listed brand. The run files record no cause.

GPT-5.2's answers named few listed brands. 24 of its 40 husband-question answers (60%) and 52 of all its 80 answers (65%) named none of the 95 listed brands. Its answers favoured gift categories over product names. Six of the 24 husband answers with no listed brand named Steam, which is outside the list. The data cover one API model on one day and give no reason for the pattern.

The three models named different brands. On the husband question the Jaccard index (brands two models both named, divided by brands either named) was 14% for Gemini and GPT-5.2, 57% for Gemini and Grok and 8% for GPT-5.2 and Grok. Across the nine question-and-model-pair cells it ranged from 4.8% (2025 question, Gemini and GPT-5.2) to 57.4%.

---

The Data That Matters

Husband question, 40 answers per model, 29 December 2025:

• Grok consistency score: 77. Coefficient of variation 23.1%; mean 10.60 listed-brand mentions per answer; range 6 to 17.

• Gemini consistency score: 48. Coefficient of variation 51.6%; mean 11.03; range 0 to 20; run means 14.0, 3.7, 13.1 and 13.3.

• GPT-5.2 consistency score: 0. Mean 0.70 mentions per answer; 24 of 40 answers named no listed brand.

• Brands in all three models' answers: 4. Apple, Sony, Bose and Patagonia, out of 61 brand names across the three models.

• Wife question against husband question: 33% to 61% fewer listed-brand mentions per answer. Gemini 6.28 against 11.03 (43% fewer), Grok 4.13 against 10.60 (61% fewer), GPT-5.2 0.47 against 0.70 (33% fewer). The Limits section explains how this gap depends on the counting method.

• Amazon in 38 of Grok's 40 husband-question answers (95%). Grok's set for that question concentrated on a few large brands.

---

The Kindle pattern

Kindle appeared in all 29 of Gemini's wife-question answers and in 7 of its 40 husband-question answers (18%). Same product, same brand, one changed word in the question.

---

The Amazon pattern

Amazon appeared in 38 of Grok's 40 husband-question answers (95%). In this test Grok's brand set for that question concentrated on a few large brands, so a steady count came with a narrow set.

---

The GPT-5.2 pattern

GPT-5.2 named none of the listed brands in 60% of its husband-question answers. A visibility check on that model and question measured a model that returned few named brands on that day.

---

Why Repeat the Measurement

One run shows one draw. Gemini's mean fell from 14.0 to 3.7 between two runs 65 minutes apart, with no change to the prompt. A single snapshot taken during either run would have told a different story.

Models differ. Gemini and GPT-5.2 shared 14% of the brands either named on the husband question; Gemini and Grok shared 57%. A result on one model says little about the others.

Questions differ. The wife question produced fewer listed-brand mentions per answer than the husband question on all three models, and Kindle moved from 18% of Gemini's husband answers to 100% of its wife answers.

This study measured which brands three models named on one day. It did not measure sales, and it does not rank vendors.

---

What You Should Do Now

1. Measure each model your buyers use

Gemini 3 Flash Preview, GPT-5.2 and Grok-4-1-fast-reasoning returned different brand sets and different counts in this test. Perplexity was not tested. A result on one model does not transfer to another.

2. Vary the question

No brand appeared in all nine combinations of question and model. 21 brand names appeared in the answers to both the husband and the wife question. Test the framings your buyers use: recipient, timing, category-level and brand-level wording.

3. Measure the spread, not one appearance

A brand that appears in 10 of 10 repeated answers stands on firmer ground than one that appears in 3 of 10. Record every repeat, read the share, and read the consistency score as a summary of how much the count varied.

4. Repeat over time

In our husband-question runs Gemini's mean brand count moved between runs within 65 minutes. A monthly snapshot can miss a shift of that size. Set the repeat interval to how often your category changes.

5. Put the repeat checks on a schedule

Choose the buyer questions that matter, ask each on every model your buyers use, and keep every answer. Rankfor runs this inside its managed AI Search programme: how Rankfor checks a brand across agreed questions and models, with repeat checks and a quarter-end review.

---

The Charts

1. Consistency score by model. Grok 77, Gemini 48, GPT-5.2 0 on the husband question. The score is 100 minus the coefficient of variation of brand mentions per answer, floored at 0.

2. Brand overlap. Three circles: Gemini (49 brand names), Grok (47) and GPT-5.2 (7), with 4 brands in the centre. Jaccard indices 0.57 (Gemini and Grok), 0.14 (Gemini and GPT-5.2) and 0.08 (GPT-5.2 and Grok).

3. Wife against husband. Mean listed-brand mentions per answer for each model under the two questions: 11.03 against 6.28 (Gemini), 10.60 against 4.13 (Grok), 0.70 against 0.47 (GPT-5.2).

4. Gemini across four runs. Run means 14.0, 3.7, 13.1 and 13.3 mentions per answer; the runs ended at 15:21, 16:26, 16:40 and 23:23 UTC. The prompt was identical in each run.

---

Data and method

• Answers: 239 (120 husband, 89 wife, 30 for the 2025 question)
• Collection: 29 December 2025, 15:16 to 23:34 UTC, through the vendors' APIs
• Models: gemini-3-flash-preview, gpt-5.2, grok-4-1-fast-reasoning, as recorded in the run files
• Repeats: rounds of 10 answers per model and question; four rounds for the husband question, three for the wife question, one for the 2025 question
• Brand count: mentions of 95 known consumer brands in 7 categories, matched against a fixed list
• Consistency score: 100 minus the coefficient of variation (population standard deviation divided by the mean) of brand mentions per answer, floored at 0
• Overlap: Jaccard index of the brand-name sets two models produced on one question
• Raw data: the eight run files, and the excluded ninth run, are in the public replication package of our gender-bias study on the same collection, Zenodo record 10.5281/zenodo.21904788 (CC BY 4.0, published 12 August 2026), folder deposit/data, files dice-roll-raw-data-<id>.json for the run ids listed under Limits

The research was conducted by the Rankfor.AI research team in December 2025. The figures were recomputed from the run files on 4 November 2026.

---

Limits of this study

• One day, three English questions. Every answer was collected on 29 December 2025. The results describe those model versions on that day under three questions. They say nothing about what the models return in November 2026 or under other questions.
• Sample. 40 answers per model on the husband question, 29 or 30 on the wife question, 10 on the 2025 question. The 2025 figures rest on 10 answers per model.
• Model versions. The run files record gemini-3-flash-preview, gpt-5.2 and grok-4-1-fast-reasoning for seven of the eight runs. The file of the run with the 3.7 mean records no model identifiers, so a version difference in that run can neither be confirmed nor ruled out. The API models ran without the consumer apps' own system layer, so these are API results, not ChatGPT, Gemini or Grok app results.
• Sampling settings. The run files record no temperature. The study notes of December 2025 give 0.7 and the later data package records the vendor defaults, so we treat the setting as not recorded.
• Counting method. The 95-brand list counts product lines such as Apple Watch and AirPods separately from Apple and ignores brands outside the list. Six of GPT-5.2's 24 husband-question answers with no listed brand named Steam.
• An excluded run. A fifth husband-question run (experiment id exp-1767026012969, 16:27 to 16:33 UTC, Gemini mean 2.5, GPT-5.2 10 of 10 answers with no listed brand, Grok mean 10.8) sits in the run files and outside the 40-answer set. No exclusion rule was recorded at the time. With it included, Gemini's husband mean is 9.32 over 50 answers and GPT-5.2's share of answers with no listed brand is 34 of 50 (68%).
• The wife-against-husband gap depends on the counting method. A later two-model extraction of the same 239 answers, published in the Zenodo package above, finds the Gemini collapse again (run means 24.9, 3.1, 24.3 and 23.0) and GPT-5.2 naming no brand in 22 of 40 husband answers. It puts the wife-against-husband gap at 4% for Gemini, 6% for Grok and 62% for GPT-5.2. The 33% to 61% figure in this article holds for the 95-brand list.
• No test of cause. The study recorded what the models returned. It did not test why Gemini's count fell or why GPT-5.2 named few brands.
• What one answer cannot establish. One answer from one model shows one draw. It cannot show whether a brand's presence is stable, how far the count moves between runs, or how the model compares with another.

Run ids: exp-1767021709939, exp-1767025618951, exp-1767026430057, exp-1767050633171 (husband), exp-1767031570132, exp-1767050965350, exp-1767051294694 (wife), exp-1767045856587 (2025).

---

Related Research

Kindle: 100% of Wife Queries. 0% of Husband Queries. Same AI.

Companion analysis of 299 answers from the same December 2025 collection, with a partner question added: how the named brands changed with the recipient's gender.

---

Key Terms

• Consistency score: how much the brand count varied across repeated answers to one question, calculated as 100 minus the coefficient of variation, floored at 0. Grok's 77 on the husband question means its counts varied less than Gemini's 48. It does not mean that 77% of answers were identical.

• Cross-model overlap (Jaccard index): brands two models both named, divided by brands either named. 14% for Gemini and GPT-5.2 on the husband question means about 1 brand in 7 was shared.

• Brand density: mean number of listed-brand mentions per answer. On the husband question Gemini averaged 11.03 and GPT-5.2 0.70.

• Run-to-run change: how much a mean moves between repeated runs of the same question. Gemini's mean fell 74% between two runs 65 minutes apart.

• Brand concentration: how far one model's answers to one question centre on a few brands. Amazon appeared in 38 of Grok's 40 husband-question answers (95%).