How AI rewrites your query before it searches
GolOps research — 37,662 prompt-to-query pairs, 664 brands. Exact word match 0.25%, fully rewritten 33.5%, average word overlap 26%. The anatomy of the invisible layer that decides who AI finds at all.
GolOps Team
Between a user's question and an AI's answer sits an invisible step. Before it finds anything, the model rewrites the query: it adds words the user never typed, changes the tone, reshapes the intent. What goes to search is not what the person wrote, but what the model decided to look for instead. No one sees this layer, yet it is the one that decides who makes the shortlist. And it is built in by design: the peer-reviewed Query Rewriting for RAG paper lays out a rewrite-retrieve-read framework in which an LLM reformulates the query first and only then goes to search.
GolOps measured it. 37,662 unique prompt-to-query pairs, 664 observed brands. The data source is the actual search queries AI systems generate when processing real prompts. Not a synthetic benchmark, but production behavior.
| Metric | Value |
|---|---|
| Prompt-to-query pairs analyzed | 37,662 |
| Brands in sample | 664 |
| Exact word match | 0.25% |
| Average word overlap | 26.2% |
Key findings
0.25% — exact match. Only one query in four hundred goes to search as exactly what the user typed. In the other 99.75%, AI searches a rewritten version.
33.5% — complete replacement. One query in three is rewritten until the original prompt and the search query share no words at all. Content that answered the user's question precisely has no chance of matching what actually went to search.
26.2% — average overlap. On average, an AI query keeps only a quarter of the words from the original prompt. Three quarters of the intent is rewritten on the way.
+42% — query expansion. The average 9.0-word prompt becomes a 12.8-word query. Each added word is one more filter your page has to pass.
How much AI changes your words
The degree of rewriting is not evenly spread. GolOps broke all 37,662 pairs down by word overlap (Jaccard similarity):
| Degree of change | Word overlap | Pairs | Share |
|---|---|---|---|
| Complete rewrite | 0–15% | 12,609 | 33.5% |
| Significant rewrite | 15–30% | 10,939 | 29.0% |
| Moderate change | 30–50% | 8,727 | 23.2% |
| Minor change | 50–75% | 4,720 | 12.5% |
| Near-exact | 75–100% | 667 | 1.8% |
The top two rows — complete and significant rewrites — make up 62.5% of the sample. Near-exact matches, where a page built around the user's wording still has a chance to land, account for 1.8%. The selection runs on the machine interpretation of the user's words, not the words themselves.
Researchers have shown that simply paraphrasing a prompt can produce up to a 100% difference in which brands get recommended — a shift invisible to the user. When AI also rewrites the query before searching, the effect on brand visibility compounds. The same spread shows up between models: Same question, different AI, different answers. Models agree 4% of the time.
The words AI adds
AI does more than rephrase the question — it injects words the user never typed at all. Google officially calls this query fan-out: AI Mode breaks a question into subtopics and issues multiple searches on the user's behalf. GolOps counted 77,068 injections across 50 unique added words. These additions decide which results appear and which brands make the shortlist.
| # | Added word | Queries | Share |
|---|---|---|---|
| 1 | brands | 8,253 | 21.9% |
| 2 | companies | 6,137 | 16.3% |
| 3 | best | 5,764 | 15.3% |
| 4 | list | 5,229 | 13.9% |
| 5 | 2025 | 4,750 | 12.6% |
| 6 | top | 4,090 | 10.9% |
| 7 | official | 3,621 | 9.6% |
| 8 | platform | 2,473 | 6.6% |
| 9 | platforms | 2,443 | 6.5% |
| 10 | software | 2,386 | 6.3% |
The leading injections are "brands", "companies", "best", "list", "top". AI assumes in advance that the user wants a ranked selection, even when the user asked a simple question. The consequence is direct: if your content does not carry words like "best", "top", and the current year, it may not match the queries AI actually sends to search — even when the user's original prompt describes your product perfectly.
Year injection
A separate mechanism is the addition of a year. GolOps recorded 6,645 year injections across 37,662 queries: AI appends the current or a recent year in 17.6% of cases.
| Year | Queries | Share of injections |
|---|---|---|
| 2025 | 4,690 | 70.6% |
| 2026 | 1,373 | 20.7% |
| 2024 | 575 | 8.7% |
A query for "best CRM software" goes to search as "best CRM software 2026" — even when the user did not ask for current results. A single year, 2025, accounts for 71% of all injections. This freshness bias puts an invisible expiry date on your content: a page without a year reference can be filtered out, even if it is genuinely evergreen.
The brands AI invents
A query for "best email marketing tool" becomes, in AI's hands, "Mailchimp vs Klaviyo email marketing comparison". The model does more than look for an answer — it inserts brand names the user never mentioned. This happens in more than one query in ten: training on comparison content leads AI to construct brand-specific searches even from a generic question.
The names inserted most often: google, microsoft, ibm, oracle, sap, intel, square, vanguard, accenture, hubspot. These are the brands that dominate comparison content in tech, SaaS, and e-commerce. They appear in AI search queries even when the original prompt contained no brand mention at all. The field of choice fills with names before the model has even started searching, and the places go to those whom comparison content has already entrenched.
Lists instead of questions
A simple question like "good project management tools" is rewritten into a format-specific search — "best project management tools list 2026". Format keywords are added to 54% of queries; GolOps identified 11 of them in total.
| # | Format keyword | Queries | Share |
|---|---|---|---|
| 1 | "best" | 5,821 | 15.5% |
| 2 | "list" | 5,567 | 14.8% |
| 3 | "top" | 4,187 | 11.1% |
| 4 | "review" | 1,784 | 4.7% |
| 5 | "reviews" | 1,300 | 3.5% |
| 6 | "alternative" | 448 | 1.2% |
| 7 | "alternatives" | 447 | 1.2% |
| 8 | "vs" | 366 | 1.0% |
| 9 | "comparison" | 196 | 0.5% |
| 10 | "guide" | 149 | 0.4% |
Three keywords — "best", "list", "top" — together appear in more than half of all format injections. This gives comparison content a structural advantage. Even when a user wants your specific product, AI searches for "best X list" — a list where you stand next to competitors, or are absent entirely.
Questions become keywords
Users ask questions. AI searches keywords. 37.8% of prompts are phrased as questions (14,236 of 37,662). In search queries, questions fall to 0.22% — 82 of 37,662. The question format is reduced by a factor of 174.
| Share | Absolute | |
|---|---|---|
| Prompts that are questions | 37.8% | 14,236 of 37,662 |
| Queries that are questions | 0.22% | 82 of 37,662 |
When a user asks "What's the best way to…", AI strips the question mark, drops the filler words, and assembles a keyword-dense search. The conversational tone disappears entirely: "What's the best project management tool for remote teams?" becomes "project management tools remote teams 2026 comparison". FAQ content written around live questions does not match what AI actually sends to search — even when the user's question describes your product precisely.
The audience AI assumes
A user asks "best CRM software". AI searches "best CRM for startups" or "best CRM software for small business" — ascribing an audience no one specified. GolOps counted 427 such audience fabrications.
| # | Segment | Queries | Share |
|---|---|---|---|
| 1 | for startups | 139 | 32.6% |
| 2 | for small business | 131 | 30.7% |
| 3 | for beginners | 84 | 19.7% |
| 4 | for enterprise | 22 | 5.2% |
| 5 | for saas | 18 | 4.2% |
| 6 | for agencies | 9 | 2.1% |
This implicit segmentation decides which content appears. If your pages target "enterprise" while queries get "for small business" injected, you become invisible — regardless of how relevant you actually are. And the reverse. The user did not pick the audience; the model did it for them.
The language switches to English
When a user asks in French, Spanish, or Japanese, AI often translates the query to English before searching. Non-English prompts make up 7.0% of the sample (2,639). In search queries, that figure drops to 2.5%.
"Meilleurs restaurants Paris pour dîner romantique" goes to search as "best romantic dinner restaurants Paris". "日本の最高のラーメン店はどこですか" becomes "best ramen shops Japan location". This creates a hidden bias toward English-language content in AI responses — regardless of where the user is or what language they asked in. For anyone serving non-English markets, the absence of English versions of key content means falling out of the choice exactly where the user asked in their own language.
The query grows longer
AI does more than rephrase — it expands. The average 9.0-word prompt becomes a 12.8-word query: +3.8 words, or +42%. 55.9% of queries expand.
| Direction | Share |
|---|---|
| Expanded | 55.9% |
| Unchanged in length | 30.2% |
| Shortened | 13.9% |
The longer the query, the higher the specificity — and the more chances your content has to miss. Each added word is one more filter the page must clear to appear in results. Expansion is four times more likely than compression.
Methodology
What underpins the numbers:
- 37,662 prompt-to-query pairs — unique pairs after deduplication, all drawn from the actual search behavior of AI systems processing real prompts. Synthetic benchmarks are not part of the sample.
- 664 observed brands — a sample across industries and sizes, from global corporations to mid-market B2B players.
- Query source — current large language models with web search, deployed in production interfaces.
- Analysis — word-level tokenization, Jaccard overlap, pattern matching (year injection, brand and audience fabrication, format keywords).
We measure not the model's theoretical capability, but the query it actually sent to search while processing a live prompt.
What this means for your company
AI almost never searches your actual words. Exact match is 0.25%, average overlap is 26%, and one query in three is rewritten to zero shared words. Between the buyer's question and the search sits a layer that appends "best", "top" and the current year, substitutes competitors' brand names, and assumes an audience no one specified. The question is rewritten on the way, and a company whose content does not match the rewritten query falls out at the first cut — no matter how relevant it actually is.
This layer cannot be optimized blind; it has to be measured — which queries AI really generates for your scenarios, which words it injects, who it puts in your place. That is the layer GolOps takes under management. We fix a company's position in the field of choice through the Choice Control Index, break down the scenarios and rewritten queries that shape it, translate the measurement into a prioritized plan, and confirm the commercial effect with a re-measurement. The Strategic Pilot closes the first cycle in 10–12 weeks; the Command Center keeps the loop running continuously across seven AI systems.
There is no time to ease into it. Gartner forecasts 90% of B2B procurement under autonomous AI agents by 2028, and Semrush already shows AI-channel conversion running 4.4× higher than organic search. The buyer sees a finished recommendation and never suspects the question was rewritten — and while AI keeps your words in only 0.25% of queries, every quarter without managing this layer costs you decisions made on a query you never saw: 99.75%.
The rewritten query is only half the story; the other half begins when an AI crawler comes for your content:
When AI comes to your website. The anatomy of 600K crawler visits