Your Performance report may already hold queries that read like someone typing to a chatbot: a platform, a timeline, two suspected causes and a decision. They sit in the long tail, where keyword tools rarely look and few people scroll by hand. A regex filter pulls them out.
This guide walks through a repeatable audit: filter, export, cluster, map each cluster to a page, and mark it covered, partial or gap. It builds on the validation step in Search Engine Journal's piece on finding the questions customers ask AI, which treats Search Console as a way to check which questions already reach your site, not as a discovery tool. So start with questions from calls, tickets and forums, then see which of them show up.
Step 1: Enter the regex patterns
Open Performance, then Search results. Choose New (or Add filter), pick Query, and change the dropdown from "Queries containing" to "Custom (regex)". Keep it on "Matches regex", paste a pattern and apply. Set the date range to the longest the report allows, since rare queries need time to accumulate.
Search Console matches a pattern anywhere in the query unless you anchor it. The patterns below use ^ (start of query) and $ (end of query) so they test the whole string.
Pattern 1: queries of eight words or more
^([^ ]+ ){7,}[^ ]+$
Read it as: seven or more chunks of non-space characters, each followed by a space, then one final word. That is eight words at minimum. Search Engine Journal's article uses this same pattern.
Pattern 2: queries that start with a question word
^(how|what|why|which|should|can|do|does|is|are|when|where|who)\s
The \s forces a space after the starter, so "cannibalization" does not match on "can". This is also the article's pattern.
Pattern 3: combined variant
To get either kind in one pass, join them with a pipe:
^(how|what|why|which|should|can|do|does|is|are|when|where|who)\s|^([^ ]+ ){7,}[^ ]+$
The pipe means "or". Everything long or question-led comes back together, which makes it the best first export. See read about how to check core update impact during holiday traffic for additional background.
How to adapt them
Change the number in the braces to move the length threshold. {5,} gives six words or more, and {9,} gives ten or more. If your site has thin data, try the lower threshold.
For questions that are also long, chain the pieces: ^(how|what|why|which|should|can|do|does|is|are|when|where|who)\s([^ ]+ ){6,}[^ ]+$ returns question-led queries of eight words or more. Add words that fit your market to the starter list, such as "will", "could" or "need". If capitalized queries slip through, put (?i) at the very front of the pattern to ignore case.
Step 2: Verify the regex returns what you meant
Do this before anyone sees a result. A wrong pattern produces clean-looking nonsense.
- Export the filtered table and add a word-count column in your spreadsheet: =LEN(TRIM(A2))-LEN(SUBSTITUTE(TRIM(A2)," ",""))+1. Pattern 1 results should never show a count below 8.
- Sort by word count and read the shortest and longest 15 rows. Look for pasted product codes, URLs or error strings, which are long but not questions.
- Switch the filter to "Doesn't match regex" with the same pattern and skim the top rows by impressions. If obvious questions appear there, the starter list is missing words.
- Compare two date ranges, for example the latest three months against the three before. A query that appears in both is a pattern. A query in only one window may be noise.
- Read the queries the way a buyer would. A query that names a platform, a constraint or a decision is meaningful. A long query that is a scraped title is not.
- Cross-check against your support tickets and call notes. If a scenario shows up in both places, it carries more weight than one that shows up in either alone.
Step 3: A worked audit (illustrative data, not real)
The dataset below is invented to show the method. It imagines a small consultancy that sells ecommerce SEO to furniture retailers, and the scenarios echo the discovery-call examples in Search Engine Journal's article. None of these queries came from a real property.
Illustrative export after applying the combined pattern, 12 rows:
- how long does it take to recover rankings after a shopify migration
- why did my rankings drop again a year after moving to shopify
- we moved to shopify and lost traffic is it the migration or something else
- how do i check redirects after replatforming a store
- do collection pages compete with each other on shopify
- can two collection pages cannibalize each other in google
- what is the difference between canonical and noindex for collection pages
- should i use separate content for france and germany sites selling the same products
- is hreflang enough for the same products in different countries
- how much does an seo audit cost for a small ecommerce store
- who offers ecommerce seo audits for a furniture store
- shopify seo agency for furniture stores selling in multiple countries
Group by the situation behind the query, not by shared words. Rows 1 to 4 describe the same moment: a Shopify move and its fallout. Rows 5 to 7 are about collection pages competing. Rows 8 and 9 are the same international decision. Rows 10 to 12 are buying-stage searches.
That gives four clusters from twelve rows. The clusters are your plan. The rows are the evidence.
Step 4: Map each cluster to a page and give a verdict
For each cluster, find the single page that should answer it. Open that page and read it against the cluster's queries.
- Covered: the page answers the situation with specifics.
- Partial: the page touches the topic but misses a sub-question.
- Gap: no page answers it.
Copy this template into a sheet:
Also read: block gptbot, allow oai-searchbot: a robots.txt audit in depth
| Query cluster | Example queries | Matching page | Verdict | Evidence | Next action | |---|---|---|---|---|---| | Post-migration recovery | Rows 1 to 4 | Shopify SEO service page | Partial | Page covers migrations and redirects, not why a second drop happens | Add a section on second-drop causes and a recovery timeline | | Collection page cannibalization | Rows 5 to 7 | None | Gap | Three distinct queries, also raised on a sales call (illustrative) | Write a guide, or add a section to the Shopify page first | | International duplicate content | Rows 8 and 9 | International SEO page | Partial | Explains hreflang, skips the "separate content?" decision | Add a decision framework with two or three examples | | Audit pricing and vendor search | Rows 10 to 12 | Services and pricing pages | Covered | Both pages state scope and price range | None now, recheck in a quarter |
The Evidence column is where your calls and tickets come in. If a cluster appears in Search Console and in three support tickets, the verdict has weight. If it appears in only one place, say so.
Search Engine Journal's article notes that strengthening an existing page is usually quicker than publishing a new one. Default the next action to that, and reserve new pages for true gaps.
For priority, weigh business value, how much evidence you have, how large the coverage hole is, and whether you can answer with real specifics. Frequency from calls and tickets stands in for the volume figure you won't have.
Step 5: State what the data cannot show
Stakeholders want a number next to every line. Give them the evidence you have and be plain about the rest. Use this checklist when presenting:
- Anonymized queries are missing. Google hides rare queries, and conversational ones are rare by nature. The Search Engine Journal article cites one analyst's finding that 57.7% of impressions sat in the anonymized pool over 59 days. Treat every result as an undercount. Treat absence as inconclusive: a missing scenario may be a gap, or it may just be too rare to show.
- There is no search volume. Impressions in your property describe your own exposure, not market demand. Never convert them into a demand estimate.
- Impressions do not prove an AI assistant was involved. A long query could be typed by a person, spoken, or generated by a search feature. Call these "conversational-style queries", not "AI queries".
- Your own calls skew toward people who already know you. Audiences you have not reached are underrepresented.
- Say what was checked. List the pattern, the date ranges, the word-count spot check and the support-ticket cross-check.
- Do not promise rankings or traffic from closing a gap. The honest claim is that you found questions customers ask and pages that do not fully answer them.
Present the output as a well-evidenced shortlist. Search Engine Journal's article recommends the same framing for stakeholders used to volume figures.
What to expect next
The first run may look thin. Rare queries are anonymized, so any list you get is an undercount. Rerun the combined pattern monthly, keep the sheet, and watch whether clusters recur or fade. Over a few rounds, the next-action column can start to read like a content calendar built on questions customers actually asked.



