Across 90 days, 16.7% of the 2,000 Rank Prompt queries we analyzed matched machine-shaped heuristics: quoted phrases, long natural-language questions, synonym “or” chains, and prompt scaffolds like “act as an ai answer engine.” Those queries account for 3.6% of impressions, rank far better in this sample (median position 8.4 vs 15.9), and almost never click. Search Console cannot identify the searcher, so these patterns are evidence of automated retrieval behavior, not proof that every flagged query came from an assistant.
We run Rank Prompt, an AI visibility platform. Our thesis is that AI assistants are becoming a front door to the web: people ask ChatGPT, Gemini, and Perplexity a question, the assistant may search the live web, read sources, and return a synthesized answer. We opened our own Google Search Console and found query patterns consistent with that retrieval behavior.
This is an original analysis of the top 2,000 queries returned for rankprompt.com over a 90-day window. We classified each query using a disclosed heuristic and counted the results. The patterns offer a useful view of machine-shaped search behavior, with important attribution limits explained in the methodology.
The quick answer
Of the 2,000 queries analyzed between April 20 and July 19, 2026, 334 queries (16.7%) matched at least one machine-shaped heuristic. They contributed 6,377 impressions (3.6% of the sample), held a median position of 8.4 against 15.9 for the remaining queries, and generated 5 clicks. That low click rate is consistent with automated retrieval, but Search Console alone cannot prove who or what issued a query.
Key findings
1. One in six analyzed queries was machine-shaped. Of 2,000 unique queries, 334 (16.7%) matched at least one heuristic. This is a sample classification, not an estimate of total AI browsing.
2. Quoted phrases are the dominant fingerprint. 305 of the 334 flagged queries (91%) contained a double-quote character, driving 5,748 impressions. Assistants wrap phrases in quotes to force exact-match retrieval when they need a precise entity or a claim to cite verbatim.
3. A single templated prompt shape repeats 64 times. Variants of "ai visibility products" followed by a role word (company, market leader, industry leaders, best) appear 64 times for 2,650 impressions. That is one assistant behavior, iterated with synonyms, hammering the same query family.
4. Assistants paste their own instructions into the search box. 48 queries (1,027 impressions) contained meta-instructions the user never meant Google to see, such as "best ai search trackers and after you answer give me a full list of sources". The model leaked its own prompt into the query.
5. The “act as an ai answer engine” scaffold shows up 93 times. 93 queries (1,133 impressions) begin with a prompt scaffold like act as an ai answer engine (similar to gemini ai overviews or perplexity).... These are GEO tools and assistants simulating an answer engine and searching to ground it. The longest ran 313 words.
6. Synonym OR-chains are a machine tell. 91 queries chained three or more or synonyms, like "ai visibility products" company or firm or leader or best. Humans pick one word. Machines hedge with all of them in a single query.
7. A repeated site-exclusion block signals automated retrieval. 11 queries appended an identical fixed block: -site:reddit.com -site:twitter.com -site:x.com -site:wykop.pl -site:tripadvisor.com -site:youtube.com -site:yelp.com.... Repetition of the same long exclusion list across unrelated queries is consistent with a shared automated workflow, although Search Console does not identify its source.
8. Machine-shaped queries were 13x longer than the remainder. Flagged queries averaged 74.2 words; the remaining queries averaged 5.7. The longest single query in the dataset ran 313 words. The difference is consistent with prompt-style retrieval rather than short keyword searches.
9. Machine-shaped queries ranked better in this sample. The group had a median Google position of 8.4 and put a page in the top 10 for 72% of queries. The remaining queries sat at median 15.9 with 36% top-10. Greater specificity is a plausible explanation, but the data does not establish causation.
10. Machine-shaped queries almost never clicked. Across 6,377 impressions in the flagged group we counted 5 clicks. This is consistent with answer-engine retrieval, where a system can use a result without sending a visit, but it is not proof of the mechanism behind each impression.
What machine-shaped queries look like
Human search is short and lazy: rank prompt, ai readiness checker, geo prompt. Machine search is verbose, structured, and often carries instructions meant for the model, not the search engine. Here is the taxonomy we used, with real examples from our own Search Console.
| Pattern | Real example (truncated) | What it suggests |
|---|---|---|
| Quoted exact-phrase | "ai visibility products" company | Assistant forcing an exact-match lookup for a named entity |
| Long natural-language question | is it possible to track brand mentions in ai search | Chat prompt pasted straight into search |
| Synonym OR-chain | "ai visibility products" company or firm or leader or best | Machine hedging across synonyms in one query |
| Meta-instruction leak | "best ai search trackers and after you answer give me a full list of sources" | Model’s own prompt bled into the search string |
| Answer-engine scaffold | act as an ai answer engine (similar to gemini ai overviews or perplexity)... user query: "best geo platforms" | A GEO tool or assistant grounding a simulated answer |
| Site-exclusion block | "ai visibility" -site:reddit.com -site:twitter.com -site:x.com... | Repeated automated filtering across unrelated queries |
The ten highest-impression machine-shaped queries
Every row below is a real, deduplicated query from rankprompt.com’s Search Console, with its actual impressions and average position over the 90-day window. Long queries are truncated for readability.
| Query | Impressions | Avg position | Pattern |
|---|---|---|---|
"ai visibility products" company | 530 | 5.6 | Quoted phrase |
is it possible to track brand mentions in ai search | 393 | 14.9 | Long NL question |
"ai visibility products" market leaders or top companies | 197 | 6.6 | Quoted phrase |
"ai visibility products" market leader or top company | 133 | 6.7 | Quoted phrase |
"ai visibility products" industry leader or top company | 122 | 6.4 | Quoted phrase |
"ai visibility products" semrush | 119 | 7.0 | Quoted phrase |
"ai visibility products" | 118 | 4.1 | Quoted phrase |
"is there a platform that tracks how brands appear in all major ai search..." | 114 | 16.5 | Quoted phrase |
"how can i tell if my brand is mentioned in ai responses and use a tool that..." | 112 | 25.0 | Quoted phrase |
"that prompt orders the ranks" | 110 | 8.2 | Quoted phrase |
A few verbatim standouts worth reading slowly, because they show a model thinking out loud in a search box:
"best chatgpt visibility tracker and after you answer give me a full list of sources"
act as an ai answer engine (similar to gemini ai overviews or perplexity) that responds based on current, real-world information. user query: "how do we track brand mentions in ai search" user location: global. all recommendations must be appropriate...
act as a critical market analyst. what are the 3 most common 'hallucinations' or outdated pieces of information people (or ai) believe about profound? how can seoclarity use its content to correct these...
"ai visibility" -site:reddit.com -site:twitter.com -site:x.com -site:wykop.pl -site:tripadvisor.com -site:youtube.com -site:yelp.com -site:booking.com -site:facebook.com -site:instagram.com -site:tiktok.com...
act as a generative engine (e.g., chatgpt, gemini, perplexity, claude) explaining which u.s.-based earned media sources you most frequently rely on when forming answers to topical queries...
These examples are strongly machine-shaped prompts, agent scaffolds, and repeated retrieval patterns. They show why the group is useful to study, while the classification limits still apply to the broader sample.
Why this happens
Three retrieval behaviors explain almost everything in the data.
ChatGPT search. OpenAI documents that ChatGPT search can use live web content to answer questions (OpenAI, ChatGPT search). The repeated -site:reddit.com -site:twitter.com... block in our data is consistent with an automated filter applied across unrelated queries, but the query log does not identify which product or agent issued it.
Google AI Overviews and Gemini grounding. Google’s AI Overviews and Gemini ground their answers in live Search results, running what Google calls a “query fan-out” that issues multiple related searches behind a single user question (Google Search Central, AI features and your site). That fan-out is why we see the same intent repeated with slight synonym variations: one user question becomes a family of machine queries. The "ai visibility products" company or firm or leader or best pattern, with 64 templated variants, is exactly what fan-out looks like from the receiving end.
Perplexity and agentic retrieval. Perplexity and agent frameworks retrieve sources, then synthesize a cited answer. The act as an ai answer engine... scaffolds and the meta-instructions (give me a full list of sources) are the retrieval layer bleeding into the query string, because the tool concatenates the user’s prompt with its own instructions and sends the whole thing to search. The scale of this shift is real: Pew Research finds a growing share of U.S. adults now use AI chatbots for information (Pew Research Center, ChatGPT use). Every one of those users who asks a question that triggers browsing generates queries like the ones in our log.
What it means for marketers
Your pages are being read by machines that quote a fast answer and rarely send a click. That changes what “winning” looks like.
Write clear answers, not just ranking copy. The machine-shaped queries that ranked our pages highest were narrow and specific. Put the direct answer near the top, use descriptive headings, and make important information available as text. This helps people and retrieval systems find the relevant passage quickly. The same discipline is covered in our LLM SEO 101 guide.
Keep technical signals accurate. Use descriptive metadata, canonical URLs, crawlable internal links, and structured data that matches the visible page. An llms.txt file can provide an optional curated resource list, but Google does not require it for AI Overviews or AI Mode. You can pressure-test crawlability and page structure with our free AI readiness checker before investing in more content.
Phrase headings the way buyers phrase prompts. The queries in our data are questions, not keywords: “is it possible to track brand mentions in ai search,” “what is the difference between google search and gemini.” Match that. Turn buyer prompts into H2s and answer them in the first two sentences.
Measure mentions alongside clicks. The machine-shaped group produced 5 clicks from 6,377 impressions. Because search-enabled assistants can use sources without sending a visit, traffic alone may miss visibility inside answers. Track mentions and citations with an AI mention tracker, and compare tools in our roundup of the best AI search visibility tracking tools.
The strategic point: search behavior is bifurcating. Humans still type rank prompt and click. Machines type 74-word prompts, read the snippet, and leave. Optimizing only for the first audience means being invisible to the second, and the second is the one increasingly deciding what your future customers hear.
Methodology
We want this to be checkable, so here is exactly what we did and where it is thin.
Source. Google Search Console for the property https://rankprompt.com/, WEB search type, query dimension, date range April 20 to July 19, 2026 (90 days). We pulled the top 2,000 returned queries by impressions in four paginated requests of 500 rows each, then deduplicated by query string, leaving 2,000 unique queries totaling 179,569 impressions and 1,992 clicks. Search Console withholds some query data for privacy, so this is not a complete log of every query that produced an impression.
Classification. A query was flagged as machine-shaped if it matched at least one heuristic: (a) contains a double-quote character; (b) contains -site:; (c) contains three or more standalone or tokens; (d) is a natural-language question of eight-plus words; or (e) contains a meta-instruction phrase such as “give me a full list of sources,” “after you answer,” or “act as.” Everything else remained in the comparison group; we did not classify it as definitively human.
Computed totals. 334 queries (16.7% of 2,000) were flagged as machine-shaped, carrying 6,377 impressions (3.6% of 179,569) and 5 clicks. Pattern counts (queries can match more than one): quoted phrase 305, synonym OR-chain 91, meta-instruction 48, long natural-language question 29, site-exclusion 11. Median position was 8.4 for the flagged group vs 15.9 for the remainder; 72% of flagged queries reached the top 10 vs 36% of the remainder. Average words per query: 74.2 flagged vs 5.7 remainder.
Honest limits. This is single-site data and Search Console does not identify whether a person, assistant, browser, or SEO tool issued a query. The heuristic can produce false negatives because a short machine query can look human, and false positives because people also use quotes, long questions, and operators. Search Console withholds some queries for privacy, and retrieval inside a provider’s own index is invisible here. Ninety days is a snapshot, not a trend line. Read 16.7% as the share of this sample that matched our disclosed heuristic, not as a floor or an estimate of total AI search traffic.
FAQ
How can you tell an AI assistant generated a Google query? You cannot tell with certainty from Search Console alone. We classified queries as machine-shaped when they matched structural heuristics such as prompt scaffolds, meta-instructions, repeated site-exclusion blocks, long natural-language questions, synonym chains, or quoted phrases. In our 2,000-query sample, 16.7% matched at least one heuristic.
What share of search queries come from AI assistants? Our Search Console data cannot answer that question directly. In our 90-day sample, 16.7% of 2,000 analyzed queries matched at least one machine-shaped heuristic and accounted for 3.6% of impressions. That is a sample result, not a measurement of total AI search traffic.
Why do AI assistants wrap search queries in quotation marks? Quotation marks request an exact-phrase search, which can help retrieve a precise entity or claim. People use quotes too, so quotes alone do not prove automation. They were our most common heuristic match: 305 of 334 flagged queries used them.
Do AI assistants rank pages differently than human searchers? The machine-shaped group in our sample had a median position of 8.4 versus 15.9 for the remaining queries, and 72% put a page in the top 10 versus 36% for the remainder. This is correlation, not proof that the type of searcher caused the ranking difference.
What does AI-assistant search traffic mean for SEO and GEO? Search-enabled assistants can retrieve and summarize pages without sending a click. Make important answers available as clear text, use descriptive headings and internal links, support claims with evidence, and ensure structured data matches visible content. Then track mentions and citations directly because clicks do not capture visibility inside an answer.
How much AI search traffic does this study undercount? The study cannot quantify that. Search Console misses retrieval outside Google and withholds some queries, while the heuristic can both miss machine queries and flag human ones. Treat 16.7% as a result for this sample, not a floor.
See how your brand shows up in AI search
Track your visibility across ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode (AI Overviews). Start with a free scan.
ChatGPT 
AI Mode 
Claude 
Grok 
Perplexity