Turn Search Console and Bing Queries Into an AI Prompt Portfolio

Search Console and Bing Webmaster Tools record the queries that surfaced your site, but a search query is not automatically a good AI monitoring prompt. Cross that demand with what you already monitor and you get a defensible shortlist, plus a baseline for the cohort you launch.

The GEO problem it solves

Sort the top queries by impressions, paste them into a prompt list, and you have renamed a keyword export. Impressions are observed exposure in one search engine: they say nothing about how often anyone asks that question to an assistant. A fragment like "crm agency france" carries intent but no decision context.

Query fan-out shows the web searches observed inside monitored answers and the prompts that triggered them; the prompt inventory shows what you already track. The shortlist becomes new intent with two independent signals behind it.

How the workflow breaks down

1. Capture the search-demand snapshot. list_brands fixes the brand, then get_search_console_metrics and get_bing_metrics return query or page rows with clicks, impressions, CTR and average position, one dimension per call. A missing integration returns connected false and no rows, so the run continues on whichever engine answers.

2. Corroborate with observed AI searches. get_query_fan_out returns the web searches detected inside monitored answers with their count, brand presence, competitor count and originating prompts. Paginate to cover the total: these are observed searches, not end-user prompt volume.

3. Subtract what you already monitor. get_prompts paginates the portfolio with exact text, status and answer count. No duplicate score exists, so the labelling is yours: exact duplicate, close variant, covered by a broader prompt, or new intent.

4. Launch the tagged cohort, then read it. Creating a prompt is not an MCP action, so you add the accepted ones under one cohort tag. After a full cycle, get_matrix at tag granularity and get_answers on that tag return mention rate, citations and the provider split, and get_prompt_answers shows the text behind whichever prompts look wrong.

What to do with the output

Three tests, applied in order. Decision context: does the query express a choice, a comparison, a best-of, an alternative? If not, reject it whatever its impressions. Fan-out overlap: does the same intent appear in the fan-out list? Demand: at least 100 impressions in a single engine, never summed across engines.

Overlap replaces the impression floor, it does not add to it. Accept decision context plus overlap first, then decision context plus the floor, and defer the rest. Sort by impressions inside each band, cut at the number of prompts you will add, and send regulated topics through a human gate.

Save the launch manifest first: exact prompt text, source queries, topic, funnel stage, cohort tag and providers, since a wording edit creates a different question. Never compare organic CTR with the mention rate of the prompt it inspired, and read prompts with no answers as missing data, not non-mentions.

The Complete Prompt

Turn my Search Console and Bing Webmaster queries into a reviewed AI prompt portfolio, then baseline the cohort I launch. Start with list_brands and ask which brand if there is more than one. Use the last 90 days unless I say otherwise.

1. Call get_search_console_metrics with dimension=query and limit 100, then dimension=page, then get_bing_metrics the same way. Keep each connected flag, date range and rows. If one engine returns connected false, continue on the other; if both do, stop and say no demand data is available. Never join query rows to page rows or add the two engines together.
2. Call get_query_fan_out for the same period, paginating until the total is covered. Keep query, count, brand presence, competitor count and originating prompts. If it returns nothing, say so and mark every candidate single-signal.
3. Call get_prompts with pagination and label each organic query against it: exact duplicate, close variant, covered by a broader prompt, or new intent. An empty portfolio makes every query new intent.
4. Cluster what remains by intent, keeping each raw query with its engine and metrics, and write one monitoring prompt per cluster.
5. Apply three tests per candidate and show all three: does it express a choice, does the intent appear in the fan-out list, does it reach 100 impressions in a single engine. Reject anything with no decision context. Rank the rest, choice plus overlap first, choice plus the impression floor second, sorted by impressions inside each band. Return a ranked table and a deferred table with reasons; if nothing survives, show the closest three and what they lack.
6. Ask me for the cohort tag and how many prompts I will add this cycle, then cut the table to that number. You cannot create prompts, so hand me the list with topic, funnel stage and cohort tag.
7. When the cohort has run a full cycle, call get_matrix with granularity=tags and find my tag's row, get_answers on it with pagination, and get_prompt_answers where the numbers surprise. Report answer count, mention rate and self-source citation rate by provider. A missing tag row or empty answer list means not yet generated, not zero visibility.