Ghost prompts are long, machine-templated queries that read like inputs to an AI system, not searches a person typed, and they now show up in Google Search Console. Across 25 websites we own and operate (consumer reviews, e-commerce, local and B2B), we found 647 templated prompt rows on 13 sites between March 1 and October 26, 2026. They earned 1,965 impressions at an average position of 7.1 and exactly 1 click. Monthly templated impressions went from 67 in June to 1,066 in October (through the 26th), and their share of all impressions rose from about 0.09% to about 0.30%.
The share is small. The pattern is not. These rows carry location preambles, answer format instructions, persona blocks and evaluation templates. They rank, they almost never get clicked, and they arrive in near-identical batches. Our working hypothesis is that they come from automated prompt tracking: software that runs fixed prompt sets through Google's AI search surfaces to monitor which brands appear. We cannot prove who sends them, and we say so throughout. What we can show is what they look like, how they behave in the data, and how to filter them out of your own reporting.
What a ghost prompt looks like.
Here are real queries from theoryroad.com's own Search Console data. We quote them as Search Console stored them (lowercase), trimming one at an ellipsis.
context: location: united states (not for language). do not include location references in your response. question: i want to collect payments in the field...which is better for client management, mycase or clio? provide a definitive answer, along with a list of pros and cons specific to their client management features for each.which is better for accounting features, mycase or clio? provide a definitive answer, along with a list of pros and cons specific to their accounting features for each.which is better for payment management, mycase or clio? provide a definitive answer, along with a list of pros and cons specific to their payment management features for each.i'm looking for legal case management software with crm features. which is better for crm capabilities, mycase or clio? provide a definitive answer, along with a list of pros and cons specific to crm features for each.
Two things stand out. First, nobody searching Google writes "do not include location references in your response." That is an instruction to a model. Second, the same sentence frame repeats with one slot swapped (client management, accounting, payment management, CRM). One preamble query appeared twice in our sample, word for word. That is how a prompt set behaves when a script runs it on a schedule.
The five signatures.
We sorted templated rows into five signatures. A row can match more than one, so the rows and impressions below do not add up exactly to the 647 and 1,965 totals.
- Location preamble. Opens with "context: location: united states (not for language). do not include location references in your response. question:" and then the actual question.
- Answer-format instruction. Asks for a specific output shape, most often "provide a definitive answer, along with a list of pros and cons."
- Persona tag. Frames the question as a role, such as "asking as: growth-stage cmo", or prepends a demographic persona block that can run to around 200 words.
- Evaluation template. Starts with "evaluate the" followed by a company and a topic ("evaluate the <company> on <topic>").
- Quoted prompt. A whole multi-sentence prompt wrapped in quotation marks, as if pasted from a spreadsheet.
| Signature | Rows | Impressions | Clicks | Sites | Avg. position |
|---|---|---|---|---|---|
| Location preamble | 135 | 514 | 0 | 4 | 6.4 |
| Answer-format instruction | 35 | 393 | 0 | 1 | 3.2 |
| Persona tag | 197 | 358 | 0 | 3 | 4.6 |
| Evaluation template | 225 | 532 | 0 | 8 | 9.4 |
| Quoted prompt | 56 | 177 | 1 | 10 | 15.4 |
| All templated (deduplicated) | 647 | 1,965 | 1 | 13 | 7.1 |
Templated rows averaged 46.4 words. The longest ran to 204 words. For comparison, the short-query bucket in this study is 1 to 3 words. Average positions are strong (3.2 for answer-format prompts), which is exactly what makes them dangerous in a report: they pull your average position up and add impressions without a single visit behind them.
Growth by month.
Volume was a trickle in spring and climbed through summer and fall. The total impression column is Search Console's report total for all 25 sites, so the share column mixes a query-level numerator with a report-level denominator and should be read as approximate.
- March: 14 templated impressions (9 rows) of about 2.5k total, from 1 site with data
- April: 26 impressions (7 rows) of about 4.3k
- May: 35 impressions (9 rows) of about 24.9k
- June: 67 impressions (22 rows) of about 73.9k, roughly 0.09%
- July: 124 impressions (68 rows) of about 125.3k
- August: 430 impressions (207 rows) of about 181.8k
- September: 203 impressions (94 rows) of about 402.6k
- October (to the 26th): 1,066 impressions (231 rows) of about 351.4k, roughly 0.30%
Long queries grew alongside them. Impressions on queries of 10 or more words rose from 547 in June to 6,167 in October. Not every long query is a ghost prompt (people do type long questions), but the two curves bend at the same time, and that is consistent with more machine traffic entering the long tail. Our portfolio also grew over this period, so part of every line here is simply more pages and more sites being indexed.
Clicks tell the story.
Click-through rate is where ghost prompts give themselves away. Across all 124,505 query rows, the sites earned 4,014 clicks on 1,166,767 impressions.
- 1 to 3 word queries: 0.33% CTR
- 10+ word queries: 0.024% CTR (4 clicks on 16,754 impressions)
- Templated prompts: 0.051% CTR (1 click on 1,965 impressions)
A query sitting at position 7 that a real person typed should earn the occasional click. Templated prompts at an average position of 7.1 earned one click in nearly 2,000 impressions. The simplest reading is that the "searcher" never looks at the results page the way a person does. It reads the answer and moves on to the next prompt in its set.
Methodology.
Source. Query-level Google Search Console performance data (query, date, clicks, impressions, average position) for 25 websites we own and operate across consumer reviews, e-commerce, local and B2B, including theoryroad.com. Window: March 1 to October 26, 2026. 124,505 query rows in total.
Word counts. We split each query on whitespace and bucketed by word count. This piece reports two buckets: short queries (1 to 3 words) and long queries (10 or more words). The long bucket held 4,690 rows and 16,754 impressions.
Signatures. We flagged a row as templated if it matched any of these patterns, applied case-insensitively to the lowercased query. These are the exact patterns used in the analysis:
- Location preamble:
context: ?location:|do not include location references - Answer-format instruction:
provide a definitive answer|along with a list of pros and cons - Persona tag:
asking as:|i am a \d\d?-\d\d|year old in the .{0,40}industry - Evaluation template:
^evaluate the .{3,80} on - Quoted prompt:
^"[^"]{20,}"(a query that opens with a quoted passage of 20 or more characters)
Limits. Search Console leaves anonymized queries out of query-level data, and Google's own help notes that filtering by query can change report totals because of truncation and that omission. So every count here is a floor, not a full census. Many of our sites are young and were indexed during the window, which inflates month-over-month growth. And this is a sample of one operator's portfolio, not the web. Your sites may see more, less or none of this.
What this means for your reporting.
If you report SEO on impressions and average position, part of what you report may be prompts that no human typed. On our sample it is a fraction of a percent of impressions. On a small site, a B2B site in a category that monitoring tools love to track, or a single page that answers a comparison question, the share can be much larger, and it lands at strong positions that flatter your average.
- Judge SEO on clicks, leads and revenue. Impressions are a leading indicator at best. Clicks, qualified leads and revenue are the score.
- Segment long queries. Report queries of 10 or more words as their own line. If their CTR is near zero, do not let them move your headline numbers.
- Do not read ghost prompts as AI citations. An impression on a templated prompt tells you a page ranked for that string. It does not tell you an AI answer cited you, or that a buyer saw it.
- Annotate the change. If impressions jump while clicks stay flat, check for templated queries before you celebrate or panic.
For the AI side of measurement, our study of AI assistant referrals across 31 websites looks at real visits from AI tools, and how to track AI search traffic and citations covers the GA4 setup.
How to filter ghost prompts out of Search Console.
The Search Console Performance report lets you filter queries with a regular expression. Google's help documents the steps and states that the RE2 syntax is used. Matching is not case-sensitive by default.
(?i)(context: ?location:|do not include location references|provide a definitive answer|list of pros and cons|asking as:|^evaluate the )The
(?i) flag is redundant in Search Console, which already ignores case, but it is valid RE2 and keeps the pattern case-insensitive if you reuse it in other tools.^(\S+\s+){9,}\S+$It matches nine or more word-plus-space groups followed by a final word. RE2 supports the
{n,} repetition it relies on. To catch long quoted prompts without catching short quoted phrase searches, use ^"(\S+\s+){9,}\S+"$.We tested all three patterns in Python against the theoryroad.com example queries above. The signature pattern matched all 12 of our sample templated queries and none of our short human control queries (for example "google ads conversion tracking not working" and a quoted three-word phrase). The 10+ word pattern matched strings of 10 and 11 words and rejected 8 and 9. The quoted pattern matched a 10-word quoted question and rejected a three-word quoted phrase. Note that the long-query pattern will also catch real people who type long questions, so use it to segment, not to delete.
What we are doing about it.
On our own sites we now report two numbers side by side: all queries, and queries that do not match the ghost prompt filter. Strategy decisions use the second. We also watch the templated rows as a signal in their own right. If a category starts getting hit with prompt sets, someone thinks AI visibility in that category is worth monitoring, and that is useful to know when planning AI search optimization and SEO work.
How to cite this study.
Theory Road, Ghost Prompts in Search Console, October 2026, theoryroad.com/insights/ghost-prompts-ai-trackers-in-search-console. Please cite the source and the window (March 1 to October 26, 2026, 25 websites) when you use the numbers.
What is a ghost prompt?
A ghost prompt is a long, machine-templated query in Search Console that reads like an instruction to an AI system rather than a human search. Typical signs are a location preamble ("do not include location references in your response"), answer-format instructions ("provide a definitive answer, along with a list of pros and cons"), persona tags ("asking as:") and evaluation templates ("evaluate the ... on ..."). In our 25-site sample they earned 1,965 impressions and 1 click.
Are ghost prompts bad for SEO?
They do not appear to hurt rankings. They distort reporting. They add impressions at strong average positions (7.1 in our sample) with almost no clicks, so they can inflate impressions and average position and drag down CTR. Judge SEO on clicks, leads and revenue.
Should I filter ghost prompts out of Search Console?
Filter them out of the numbers you use for decisions, but keep a view of them. In the Performance report, click + Add filter, choose Queries, choose Custom (regex), pick Doesn't match regex and paste the signature pattern from this article. Search Console uses RE2 syntax.
Do ghost prompts mean AI Overviews cite my site?
No. An impression on a templated prompt means your page ranked for that string in Search Console data. It does not show that an AI answer cited you or that a person saw it. Our hypothesis is that these come from automated prompt tracking, which we cannot confirm because Search Console does not report who sent a query.
How do I check my own site?
Apply the signature regex in the Performance report with Matches regex, then compare impressions, clicks and position against Doesn't match regex. Add the 10+ word pattern as a second segment. Or export your queries and paste them into the Ghost Prompt Detector, which runs in your browser.