The prompt you track is not the query AI actually searched

The per-question score in an AI search visibility report is not the score of that one question string; it is the result of the set of sub-queries the engine expanded the question into through query fan-out. The engine searches with sub-queries that add a year or switch to English, and how much each branch contributed to the answer is not visible from outside. So if you read the scores you get for 30 tracked questions as scores for those 30 strings alone, you can attribute the results to the wrong cause.

First, the scope and grade of our evidence. As of October 2026 we compared five public sources: one social post, three vendor analyses, and one industry article covering one of those analyses ^{[1][2][3][4][5]}. All of them are observations that were not peer reviewed, so on our evidence scale they rank medium or below, and we did not reproduce any of the numbers. The vendor analyses come from a company that sells AI visibility tracking, and we also run an AI search visibility diagnosis product, so there are interests and bias from the same market on both sides. This post uses these numbers only to set the direction of measurement design, not as grounds for scoring weights.

The prompt you track is not the query the engine searched

The tracking unit is the question, but the retrieval unit is the sub-query. A visibility report usually works by entering questions and getting scores per engine, but the engine does not search the question as-is. It expands it into several branches. That process is query fan-out, and we covered how to observe it directly in What is query fan-out? Watch AI split one question.

How far and in what shape the question expands is not fixed. In the Peec AI blog's analysis of 5 million fan-outs, ChatGPT generated 2.1 sub-queries per prompt on average, Perplexity 1.4, and Grok 6.8 ^[4]. In a separate analysis from the same company, the average length of ChatGPT sub-queries doubled in four months, from about 6 words in October 2025 to about 12 words in January 2026. The number of sub-queries per prompt stayed nearly the same over that period, at 2.3–2.8 ^[5]. Even for the same question, the strings the engine actually searches change over time.

Figure 1. The tracked question q passes through a fan-out F that cannot be seen from outside and expands into several sub-queries, and retrieval and citation happen on those branches. The figures on the right come from separate public observations, and the example queries are illustrative.

Two public observations: year injection and English sub-searches

There are two lines of public observation showing that fan-out adds strings the original question did not contain. One is the year, the other is language.

Table 1. Public observations showing that fan-out adds strings the original question did not contain ^[1][2][4].

The year observation comes from a public LinkedIn post by Malte Landwehr ^[1]. It covers more than 8,000 automated conversations built from trending stories on the Google News home and topic pages, re-asked every 6 hours until the 24-hour news cycle ended. For news prompts, ChatGPT ran a web search 100% of the time, and 94% of the follow-up queries contained "2026." Only 1% of the original prompts included a year.

This number has to be read by prompt type. In an analysis of 5 million general customer prompts, ChatGPT added the current year 5.44% of the time ^[4]. The picture is that for questions where timing is part of the answer, like news, year injection is close to the default, and for other questions it is rare.

The English observation is an analysis Peec AI Research published in February 2026 ^[2]. 78% of non-English sessions included at least one English sub-search, and by language the share ranged from a low of 66% for Spanish to a high of 94% for Turkish. The Search Engine Journal article covering this result noted that prompt composition and representativeness were not disclosed, and that the data came from the vendor's measurement platform rather than from consumer sessions ^[3]. Neither source gives a figure for Korean.

Beyond years and languages, the engine inserts words

The strings added to sub-queries are not limited to years and languages. In the analysis of 5 million fan-outs, the words added most often were "best," "top," "comparison," "reviews," "tools," and "software" ^[4]. For advice-seeking questions, "best" appeared in 24.3% of sub-queries. Even if a user only asks for "AI visibility measurement tools," the string the engine actually searches may look closer to "best AI visibility tools comparison."

Sub-queries that target specific sources were observed too. In the same analysis, Grok used Reddit in its searches in 10.5% of all conversations, and 9 out of 10 of those were searches that named Reddit explicitly with the site: operator ^[4]. In cases like this, the candidate document pool is assembled within a source range the engine chose, not from the string the user typed.

Table 2. Types of strings inserted into sub-queries and their effect on measurement. The figures come from separate observations ^[1][2][4].

None of these four appear in the question string in your report. Even with the same tracked prompts, when the strings the engine inserts change, the competing document pool changes, and the score moves with it.

A report score is V(F(q)), not V(q)

The value printed as the score for question q is really a weighted sum of sub-query scores. If we call the tracked question $q$ and write that the engine applies an invisible transformation $F$ to produce a set of queries, it looks like this.

$V(q_i)$ is how well our page is retrieved and cited for one sub-query, and $w_i$ is that branch's contribution to the final answer. Retrieval and citation happen on $q_i$, not on $q$. The problem is that neither $F$ nor $w$ can be observed from outside. What we measure is $V(F(q))$ wearing the name $V(q)$.

Two consequences follow. One is that attribution can go wrong, and the other is that the language boundary is thinner than it looks. The next two sections take them in turn.

The attribution of the year effect changes

The year injection observation adds a candidate mechanism to the claim that "putting this year in your title raises citations." We covered how that number cannot separate a freshness signal from the effect of the title string itself in Does putting the year in the title increase AI citations?. The 94% observation adds a third explanation to that gap ^[1]. If users rarely type a year but the engine adds one to its sub-queries, then a year in your title is matching a string the engine inserted rather than one the user typed.

This is not proof. It is a different observation, on news topics, and correlational. What it does give us is testable predictions.

1. The year effect should be larger for question types that get more year injection. It should be large for news and latest-comparison questions and small for definition questions.
2. It should bend once when the year turns over. When the token the engine injects changes from "2026" to "2027," the effect for pages with "2026" in the title should drop at that point.
3. If it doesn't bend, the string-matching explanation weakens. In that case the freshness-signal explanation remains.

In practice, whether to put a year in a title is a call to make by question type. If you stamp a year on a post you won't update, you also take on the risk of mismatching the string the engine injects once the year turns over.

English branches get mixed into scores tracked in Korean

A score tracked with Korean questions is likely not the result of Korean searches alone. When one sub-query goes out in English, part of the candidate document pool is assembled from English documents. A brand that targets only the domestic market is competing for only some of the slots to begin with.

Here is the counterargument too. One sub-query being in English does not mean the final answer is filled with English sources. The figure that 43% of sub-queries are in English also means the other 57% stay in the original language ^[2]. Results vary by query type, by whether the topic is local or general, and by engine, and Korean does not appear as a figure in the observed samples. So the conclusion from this observation does not go as far as "you need to publish English content." It stops at this: an English branch's share may be mixed into your Korean scores, so look at that share separately.

What you can do from outside: measure the shadow, design for sets, record the version

You cannot see fan-out $F$, but you can see the marks $F$ leaves on the output. Some surfaces have been reported to expose internal search queries, but that is not a stable observation path. Instead, here are three things you can do now.

Table 3. Measurement add-ons you can apply from outside when you cannot see fan-out directly.

First, record the language and domain mix of cited URLs. If a Korean question comes back with citations mostly from non-Korean domains, that is the mark the invisible English branch left on the output. It is a proxy, and it should only be used as one.

Second, change how you build prompt sets. Collecting questions with different strings and collecting questions that induce different sub-query sets are different jobs. The first is easy, and the second is real coverage. "AI search visibility measurement tool" and "AI visibility tracking tool recommendations" differ as strings but are likely to expand into similar sets, so adding both barely increases coverage. We covered the basics of designing monitoring prompts in Prompt tracking: How to monitor AI search citations, and how to reflect sub-queries in your content in Query fan-out GEO: 5 steps to put subqueries in titles.

Third, record on the premise that $F$ has versions. The same question expands into different sets over time. In an analysis of more than 20 million fan-outs across five countries, the average sub-query length grew from about 6 words in October 2025 to about 12 words in January 2026, peaking at about 16 words in one week in between ^[5]. That analysis did not tie the change to a specific model version. Still, when the strings the engine searches change this much while your content stays the same, the candidate pool your page competes in changes too. If the engine version is not stored alongside your measurements, you cannot separate your content's share of a score change from the ruler's share.

What this post doesn't cover

Every number in this post is a public observation, and none were peer reviewed. The 94% year figure comes from a single social post, and since we did not preserve the individual post URL, the citation points to the author's profile ^[1]. The English sub-search figures come from a vendor platform running customer-defined prompts through browser automation, so there is no basis for treating them as representative of ordinary user sessions ^[3].

The samples are skewed too. The year observation is limited to news topics, and the language observation includes no figure for Korean. We did not reproduce any of the numbers, and we have not checked whether Korean questions show the same rates. The weighted sum in equation (1) is a model that describes the measurement structure, not a claim that engines actually build answers through this kind of linear combination.

It is safer to read the numbers as a direction than to carry them over as constants. The direction itself is clear. The score you get for a question is less the score of that question than the score of a set that passed through an invisible transformation.