Not sure where to start? Take a guided tour.
AI Site Grader
AI Query Scanner
llms.txt Generator
In short
Five findings, strongest first- 0.55%
of searches matched the prompt word for word, about 1 in 180. AI rewrites almost every question before it searches.
Chapter 01 - 72%
of searches shared under half their words with the prompt. On average, 38.1% of words overlap.
Chapter 01 - 22.8%
of searches added a place the prompt never named, such as the UK or the United States.
Chapter 07 - 1 in 5
searches named a brand the prompt never mentioned.
Chapter 04 - 14.3%
of searches added the word "best", the most common format word.
Chapter 05
AI rewrites almost every question.
When you ask ChatGPT, Perplexity or another AI assistant a question, something happens before you get an answer: the AI searches the web. But it doesn’t search your exact words.
Instead, it rewrites your prompt into a search query, adding context, injecting keywords and reshaping your intent. This invisible layer decides which results the model reads and which brands get discovered.
We first measured this in January 2026 on 11,521 pairs from 100 brands. This rerun covers 112,119 pairs from 825 brands, captured between 1 September and 15 October 2026. Prompt, query and brand-level rows are withheld.
The core finding holds: only 0.55% of searches matched the prompt word for word, and 72% shared under half their words with it. Some earlier findings did not hold, and the notes say which.
The words AI adds.
The vocabulary layer
AI doesn't just rephrase your question. It adds words you never used. These additions shape which results appear and which brands get recommended.
The most common addition is "brands", added to 14.1% of searches. "best" and "official" follow close behind.
If your content doesn't use words like "brands" or "best", it may not match the searches AI actually sends, even when the user's prompt describes your product perfectly.
About 1 in 10 searches get a year.
Freshness bias
When you ask about "best CRM software", AI often searches for "best CRM software 2026" - even when you did not ask for current results.
AI assumes you want fresh information, so it automatically injects the current year (or recent years) into your queries. This freshness bias shapes which content appears in responses and creates an invisible expiration date on your content.
Content without year references may be filtered out, even if it's evergreen. Regularly updating your content with current year mentions can improve your visibility in AI-powered search.
AI adds brands nobody asked about.
Phantom competitors
When you ask "what's the best email marketing tool", AI doesn't just find answers - it inserts brand names you never mentioned. Your search for "email marketing software" becomes "Mailchimp vs Klaviyo email marketing comparison".
In the rerun this happened in about 1 in 5 searches. AI's training on comparison content leads it to assume brand-specific searches, even when you asked a generic question.
Brand insertions were counted only when the original user prompt contained no brand mention.
Questions become content types.
The shortlist bias
When you ask a simple question like "good project management tools", AI doesn't search those words. Instead, it transforms your query into format-specific searches like "best project management tools 2026".
The most added format words are "best", "review" and "reviews". "best" alone is added to 14.3% of searches.
AI's format injection creates a structural advantage for comparison content. Even if a user wants your specific product, AI searches for "best X", where your competitors appear alongside you, or without you entirely.
Questions stop being questions.
From questions to keywords
Users ask questions. AI searches keywords. This fundamental mismatch means your FAQ-style content may never match what AI actually queries - even when a user's question perfectly describes your product.
When you ask "What's the best way to…", AI strips the question mark, removes filler words, and constructs a keyword-dense search. The conversational tone disappears entirely.
If your content is written to answer "How do I…" style questions, you may miss the keyword-based queries AI is actually sending to search engines. Consider including both conversational and keyword-rich variations.
AI adds a place you never named.
In 22.8% of searches, AI added a place the prompt never named: a country, region or city such as the UK, the United States, Australia, India or London. The user asked a general question. The search went local.
That changes which pages get read. A search for “best accounting software UK” would surface UK review sites and UK vendors first, whether or not the user is in the UK. If your pages never say where you sell, they can miss these searches.
This is a floor. We matched a list of about 35 places, so searches that added a place outside it are not counted.
Some searches switch language.
English bias in search
When users ask questions in French, Spanish, Japanese, or other languages, AI often translates them to English before searching - even when the user might prefer results in their native language.
This creates an invisible bias toward English-language content in AI responses, regardless of where the user is located or what language they used.
If you serve non-English markets, having English versions of key content may improve your visibility in AI responses - even when users ask in their local language.
Searches run slightly longer.
The expansion effect
AI doesn't just rephrase - it expands. Your concise questions become longer, more specific search queries packed with additional context, years, and qualifying terms.
The average prompt of 9.8 words becomes an 11.4-word query - a 16% expansion that fundamentally changes what gets found.
Longer queries mean more specificity - and more chances for your content to mismatch. Each added word is another filter your pages must pass to appear in results.
How we did it.
Unique prompt-query pairs after deduplication
Diverse sample across industries and sizes
Mean word-level similarity between pairs
Searches sharing under half their words with the prompt
Why this data is unique
This research is built on actual AI search behavior - the queries AI systems generate when retrieving information for real user prompts. Unlike synthetic benchmarks, every data point represents production AI behavior observed across 825 brands tracked by Trakkr.
The rerun reads true prompt and search pairs from row-level citation records, 1 September to 15 October 2026, deduplicated on lower-cased text, cited rows only. Results are pooled across AIs. A brand counts as added when one of the pair’s tracked brand or competitor names appears, capitalised, in the search but not the prompt; a spot check found most hits were real brands, so we report about 1 in 5. Places use a list of about 35 countries, regions and cities, so 22.8% is a floor. The first snapshot (11,521 pairs from 100 brands, January 2026) joined prompts and searches through the cited page, which mixed up some pairs. Chapters 08 and 09 still use it.
We capture the invisible translation layer between what users ask and what AI searches. This ground-truth data reveals patterns that theoretical models miss: year injection, brand hallucination, intent escalation, and the systematic rewriting of natural language into search-optimized queries.
