What Does a GEO Score Measure? Prompts, Models, and Judges Change the Number (2026)

Key takeaway: Before comparing GEO scores, check the conditions that produced them. A methods paper published on September 6, 2026 (Martinez, arXiv:2609.06811, preprint) lays out how a GEO visibility score is built from prompt sets, prompt wording, execution conditions, weights, and scoring rules, and notes that changing only the judge instruction can change the score of an identical answer. When RanketAI asked the same set of questions on the same day and changed only the model, our own brand mention rate came out at 43% versus 5% in ChatGPT and 74% versus 40% in Gemini. With GPT-6 just released, here is how to check the conditions behind a score.

---

Three-line summary

• A GEO score is not a single fact about AI answers; it is the output of a measurement design. Which questions were collected, which model was asked and when, and what counted as a mention or citation all feed into the number.
• In our own test, changing nothing but the model moved the mention rate several-fold. So the measurement model should be judged not by how new it is but by whether it is the model the users you want to represent actually see. GPT-6 Sol and Luna launched on September 22, 2026, yet on September 24 the default model in free users' regular chat was GPT-5.6 Luna.
• The paper is a methods survey with no new experiments, so its numbers are not evidence. What it offers is a framework for what to ask when you evaluate an AI visibility tool or report.

Same questions, same day — only the model changed, and the mention rate moved

We checked how much measurement conditions move a score in our own data first. On September 4, 2026, while reviewing our measurement conditions, RanketAI ran a set of category questions that did not include any brand name on the same day, changing only the model, over repeated rounds. The target was our own brand, in the AI search visibility diagnostics category.

AI service	Model condition	Our brand mention rate
ChatGPT	Model A	43%
ChatGPT	Model B (lighter model)	5%
Gemini	Version A	74%
Gemini	Version B (newer version)	40%

In ChatGPT, the gap moved together with search depth. Model A averaged 4.3 web searches per answer and model B 1.2, and every answer that mentioned our brand also cited one of our pages. One reasonable reading is that deeper search checks more sources and widens the candidate list accordingly.

The Gemini result makes a different point. Moving to a newer version does not guarantee a higher score either. In this test, the earlier version produced the higher mention rate. Both numbers are real AI answers, so the table alone cannot say which one is "right" — what the measurement is meant to represent has to be settled first.

The limits are clear. One brand (ours), one category, a small sample, one day. The magnitudes should not be generalized to other brands or categories; what the table shows is direction, not size: the choice of measurement model alone can move the same brand's score several-fold.

The five components of a GEO score, according to the paper

The methods paper published on September 6, 2026 turns this phenomenon into a framework. Its author, Olivier Martinez, is the researcher who in July published a critical survey grading 45 GEO studies by strength of evidence. This paper steps back from "do GEO techniques work?" to ask "what does a GEO score measure?"

The paper splits the instrument that produces a GEO visibility score into five components.

Component	What it sets	What changes when it changes
Situation annotation	The usage situations the questions represent, and the rules for classifying them	The same score means something different depending on which situations were measured
Prompt formulations	How each situation is phrased as a question	Wording alters retrieval, the competing sources, and the answer itself
Execution conditions	Model and version, date and time, interface, prior conversation, parameters	The same question produces a different answer when the model changes (see our test above)
Weights	The relative importance of each situation	Weights can flip the ranking of two brands
Scoring rules	How mentions and citations are identified in the answer; for an LLM judge, its model, rubric, and instruction	The same answer can score differently when the judge instruction changes

The point of the table is simple: a different score does not by itself mean different visibility. If even one of the five components differs, the two scores may be measuring different things.

The "answer market" — your question set defines what you measure

The paper's central concept is the "answer market." The questions used for measurement, together with their weights, make up the market being measured.

"We call this weighted set of answer opportunities an “answer market.” This constructed market does not necessarily represent actual user demand." — Martinez, Measuring GEO Visibility, arXiv:2609.06811 (2026-09-06)

In the same passage, the author argues that a sentence like "a source is cited in 40% of answers" is not enough on its own. To interpret it, you also need to know whether those answers responded to open-ended requests, to prompts naming the source, or to comparison requests — and which engine and period they came from.

Two problems covered earlier on this blog fit under this one concept. The brand-name prompt post showed that prompts containing the brand name almost always produce a mention in web-search-enabled AI, so they carry no information as a performance metric. The demand-side data post showed that there is no data to check how close measurement questions are to real purchase-evaluation questions. Both are cases where the answer market drifts away from real demand.

Execution conditions — a model's launch date is not its default date

The model is the execution condition that changes most often. OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. The announcement makes clear, however, that a launch and a change of default model are not the same event.

"GPT‑6 Sol and Luna roll out today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users." — OpenAI Developer Community announcement (2026-09-22)

Surface (checked 2026-09-24)	Model provided
Default model in ChatGPT free users' regular chat	GPT-5.6 Luna
Free and Go users, desktop app trial	GPT-6 Luna
Paid users (Plus, Pro, Business, Enterprise, Edu) in ChatGPT Work and Codex	GPT-6 Sol, GPT-6 Luna

GPT-5.6 Luna's availability to free users comes from OpenAI's announcement; the regular-chat default as of September 24 was checked directly in a free account. We will update this table as models change.

When choosing a measurement model, decide first whose answers you want to represent. If the measurement is meant to represent what free users see, the GPT-6 launch itself is no reason to change conditions; the point to change is the day the default model in free users' regular chat changes. If it is meant to represent users in paid work environments, the reference point is different. Either way, what to check is whether the tool has defined its reference point and whether it records when that reference point changed.

The paper is specific about how to read scores after a model change, too.

"A stable identifier does not guarantee the absence of silent updates: interleave conditions, use sentinel controls, and segment collection if a break is suspected." — Martinez, arXiv:2609.06811, measurement protocol step 4

In practice, scores from before and after a default-model change are better not read as one continuous trend line. Record when the change happened and treat the next few rounds as a new baseline; only then can before and after be compared.

Date is an execution condition as well, because answers differ from run to run even on the same model. In our re-measurement benchmark, measuring 12 Korean B2B SaaS companies twice on the same day produced a median individual score change of 6 points, while the median change over a 20-day interval was 5.5 points — the month's change sat inside same-day re-measurement noise. How many repetitions it takes before a result can be trusted is covered in our St. Gallen study analysis.

Scoring — the same answer can score differently under a different judge instruction

Receiving the answer does not settle the score yet. A scoring step still has to identify brand mentions and citations in the answer, and if that step is handed to a language model, another instruction enters the measurement.

"This instruction can change the scoring even when the evaluated answers remain identical." — Martinez, arXiv:2609.06811

The paper proposes three safeguards: if an LLM judge is used, declare its model, rubric, and instruction; keep the raw answer text and the coding rule; and cross-check by applying a different judge to the same saved answers. From a user's side, this reduces to one question: when a score changes, can you tell whether the answer changed or the judging changed? That requires being able to see the raw answers alongside the results.

Weights — reporting a range instead of a single score

A composite score hides situation weights. The paper gives an example where the choice of weight alone flips a ranking. The author labels it a fictional example; discovery situations are questions where users look for candidates for the first time, and verification situations are questions checking something they already know.

	Discovery success rate	Verification success rate	Discovery weight 0.2	Discovery weight 0.8
System A	0.30	0.90	0.78	0.42
System B	0.50	0.60	0.58	0.52

A leads when the discovery weight is below 0.6, and B leads above it. If the weight cannot be pinned down anywhere between 0.2 and 0.8, the paper recommends reporting the score as a range: 0.42–0.78 for A and 0.52–0.58 for B. A single score hides the fact that A swings widely with the weight while B stays stable.

The practical question is whether the composite score's weights match your goal. If the goal is winning new customers, discovery situations — where users who do not know your brand look for candidates — should carry more weight; if the goal is retaining existing customers, verification situations can reasonably weigh more.

Being cited is not the same as contributing

"Our citation count went up" and "that source contributed to the answer" are different claims.

"The presence of a citation is insufficient to establish a source’s contribution to an answer." — Martinez, arXiv:2609.06811

To examine contribution, the paper proposes fixing the question, the conditions, and the list of retrieved sources, removing one source, regenerating the answer, and comparing the two. In that design, "was it cited?" cannot be the outcome measure, because removing a source mechanically removes its citation. What has to be compared is whether the information in the answer changed.

This distinction points the same way as two earlier observations on this blog. The ghost citations post covered cases where a page is cited but the brand is never mentioned, and the critical survey analysis covered how being selected as answer material and being absorbed into the answer body are separate problems.

Seven things to check when evaluating a measurement tool or report

Turning the paper's seven-step measurement protocol into questions a user can ask gives the list below. Question-design criteria were covered in the brand-name prompt post, so this list focuses on the other conditions.

Question to ask	Why it matters
Which situations does the question set represent? Does it include prompts containing the brand name?	The question set is the market being measured
Does the measurement model match the users you want to represent (free or paid)?	Changing only the model moved the mention rate several-fold
Is the date of a default-model change recorded?	Scores before and after a change cannot be read as one trend line
How many times, and over how long, was each question repeated?	Same-day re-measurement noise was about as large as a month's change
Is the mention and citation scoring method documented and held fixed?	Changing only the judge instruction can change the score
Do the composite score's weights match your goal?	Weights can flip rankings
Is a rising citation count being read as performance?	A citation is not evidence of contribution

RanketAI's tools follow the same order. A page structure check first confirms that AI crawlers can read your pages, and AI brand visibility analysis then measures mentions and citations in ChatGPT, Perplexity, and Gemini answers to questions that do not include your brand name. To see competing brands side by side under the same conditions, use competitor comparison. Whatever tool you use, the seven questions above apply the same way.

Three lines to hold when citing this paper

The paper's value lies in its framework for measurement design, not in numbers. Cite it with these limits attached.

1. There are no new experiments. The author states: "No new experimental collection from users or generative systems is reported." The ranking-reversal example is explicitly labeled fictional. Quoting its numbers as real measurement results would be a factual error.
2. It is a single-author preprint that has not passed peer review. The empirical validity of the proposed protocol is also left for future work.
3. It is a selective, predominantly English-language survey. The author notes that the same vocabulary appearing across studies is not an accumulation of independent replications. There is no evidence yet that it transfers unchanged to Korean-language questions or Korean AI services.

Our own measurement deserves the same reading. The table above comes from one brand, one category, a small sample, and one day; it shows the direction in which the measurement model moves a score, not which model is better.

FAQ

Related reading

• The 40% GEO Gain Has a Condition — Inside the 2026 Critical Survey of 45 Studies — the same author's earlier survey, on the conditions behind technique effects.
• Why Brand-Name Prompts Can't Measure AI Visibility — criteria for designing the question set.
• Monthly Change Was Smaller Than Same-Day Measurement Noise — measured variance from date as an execution condition.
• Don't Measure AI Visibility Once: Why You Need 7 Runs a Day and 2–4 Week Windows — the statistical baseline for repeated measurement.
• The Demand Side AI Visibility Misses — another side of the answer market drifting from real demand.