The Search Console AI filter is a typing shortcut, not an analyst: it turns a plain-English request into the query filters you'd otherwise click together by hand. That's genuinely useful, because the filters most worth building are the ones almost nobody builds. It's also capped, occasionally wrong, and limited to configuring what you could configure anyway. Know both halves and it earns a place in your monthly routine.
Google confirmed the full rollout in February 2026, after two months of limited testing. Here's what it does, where it breaks, and the six prompts that make your query table say something useful.
What it does, and the three limits
Type a request into the box in the Performance report ("Search results" view) and the feature converts it into filters, comparisons, and metric selections. Filter by query text, page, country, device, or date range, in your own words.
The limits, all three per Google's own rollout notes:
It's capped at 20 requests a day. Enough for a monthly audit, tight for a day of exploratory iteration.
It only configures. No sorting, no exports, no analysis. The AI box sets the table up; reading the table is still your job.
It can misread you. Google recommends reviewing the filters it applies before trusting the data, and that review takes five seconds: the filter chips appear above the table. Check them every time. The AI box saves typing, not thinking.
The six prompts
Every query table hides distinct groups of searchers, and each group deserves a different response from you. These six prompts split them apart. Run them in the Performance report; each takes seconds.
1. Show queries that contain buy, price, cost, hire, near me, or service
2. Show queries that contain vs, versus, or difference
3. Show queries that contain near me, near, or in
4. Show queries with six or more words
5. Show queries that contain fix, problem, issue, or not working
6. Show queries that contain how, what, or why for pages that contain /blog
What each surfaces, and where to route it:
Prompt 1 finds ready-to-buy searches ("seo services near me," "how much does a kitchen remodel cost"). These convert far better than anything else in the table. Route them to service and pricing pages, and work them first.
Prompt 2 finds comparison searches ("wix vs wordpress"). Decision-stage traffic that's explicitly asking to be helped choosing. Comparison pages rank because the format matches the search, and AI answers quote them heavily.
Prompt 3 finds local intent ("dentist in jacksonville"). Feed these into location pages and your Google Business Profile categories. Expect noise from the word "in" and skim before acting.
Prompt 4 finds long-tail questions ("why is my website not ranking on google"). Lower volume, higher specificity, usually lower competition: this is your content-idea backlog, pre-validated by your own data.
Prompt 5 finds problem searches ("google ads not working"). Solution-seeking traffic converts well, and help content builds trust before the sale.
Prompt 6 finds the informational questions already driving your blog. Your expand-and-update list, and your featured-snippet and AI-answer targets.
The two-step move that makes any filter pay
A filter is a question you ask your own data. The answer that matters comes from the same two steps every time:
Sort the filtered table by impressions, descending. Then flag every query with high impressions and low clicks. Those are searches where Google already shows you and searchers already skip you, which means a title rewrite or a dedicated page section is often all that stands between you and the traffic. No new content, no new links.
That flagged list is precisely the input to the striking distance workflow, and we've written the full striking distance method separately: the position band that pays, the two failure modes, and the real odds.
What this looks like on a real table
A worked example, shaped like what we see in client accounts. A Jacksonville home-services site runs prompt 1 and gets 22 commercial queries back. Sorted by impressions, the top row reads:
Query: "emergency plumber jacksonville"
Impressions: 4,100 Clicks: 12 Position: 9.3
Four thousand people saw the listing; twelve clicked. Position 9 means Google already considers the page relevant, so this isn't a content-creation project. It's a snippet problem: the title says "Services | Smith Plumbing" and says nothing about emergencies. A title rewrite naming the actual search ("24-Hour Emergency Plumber in Jacksonville") plus one H2 answering it directly is an afternoon of work against a query with proven commercial demand.
That's the pattern the six prompts exist to surface: the filter finds the group, the impressions sort finds the biggest gap, and the fix is usually smaller than you'd guess. Log every edit with the date, then recheck at 30 days, because about half of these bets move and you want to know which half.
When the box misreads you
A request like "show my best queries" is ambiguous (best by clicks? impressions? position?) and the AI will resolve the ambiguity somehow, silently. The fix is to phrase prompts the way the six above are phrased: name the exact words or the exact condition, never a judgment call. And when a filter matters, know its manual twin. Every prompt above has a regex equivalent that runs in any Search Console account with no daily cap and no interpretation step; we keep both versions side by side in the free GSC Intent Prompt Pack, copy-paste ready, along with which page type to route each intent to.
(If you're setting Search Console up for the first time, or your business doesn't run on Google Workspace, Search Console works without a Gmail account; do that first, since none of this exists without the data.)
Who should skip this for now
A site in its first few months mostly can't use any of it: the filters will return thin tables because Search Console barely has data to filter yet. That's the data being young, not the method failing. Build pages first and come back around month three. And if your curiosity runs toward how AI assistants see your site rather than how Google does, measuring your AI traffic is the parallel exercise.
For everyone with a year of data: six prompts, two sorting steps, five minutes a month. We run this exact loop as the opening move of the Assay phase in the Reforge Method, because the cheapest wins in SEO are the ones your own data has been trying to show you all along.