Which of 15 tools does one consultancy's 60-skill library use? Search Console leads; Semrush is named most

Key findings

• The sample: one SEO and AEO consultancy's working library (ours): 60 skills, holding 733 files besides their 461 recorded learnings. We measured it on 10 November 2026 (UTC) from a copy scrubbed of client names; the newest file in the copy was last modified on 5 November 2026. Every figure describes this one library, not consultants in general.
• Of the 15 outside tools we read, Search Console is the one a step uses in the most skills: 23 of the 60 (22 in a check with fuller excerpts; see below). In 18 of them, the step names it with no alternative and no condition. Semrush is named in the most skills (38 of 60), but a step uses it in 16 (16 in the check). Screaming Frog follows with 12 (12). The top three are the same in both readings. "Uses" here is what the skill's text says a step does, as AI readers labelled it; we didn't run the skills.
• Most pattern-matched pairs are mentions. The 15 tools' patterns matched 335 skill-and-tool pairs. In 96 of the 335 a step uses the tool (95 in the check); in 203 the skill only mentions it. Ahrefs is named in 22 skills and a step uses it in one.
• The answer engines are mostly named in passing. The patterns for ChatGPT, Gemini, Perplexity and Google AI Overviews matched 109 pairs: 66 are mentions, in 24 the engine is what the skill studies (where a brand shows up in its answers), in 11 a step uses the engine itself, and 8 are other (6 where the two runs disagreed, 2 where the word meant something else).
• Our MCP server's tools, as names in the text (exploratory, not read): the skills are written for an MCP server (ours) with 90 tools. 57 of the 60 skills name a tool of their own learning loop (6 tools), and 23 name none of the other 84. Page scraping and crawling is the job named in the most skills (26), then keyword and trend research (21) and search results (17).

What this doesn't show

• One library, an internal case study. One practitioner's 60 skills say nothing about which tools consultants use in general, and the library itself can't be shared, so readers can't re-run the numbers.
• Text, not runs. "A step uses it" means the skill's text has a step that uses the tool. We didn't run the skills, count how often a step ran, or check how a step gets its data.
• AI readers, not people. Two runs of one AI model labelled every pair and agreed on 326 of 335. That is consistency, not correctness: no person checked the labels.
• Fifteen tools, chosen by us, plus one partner platform we left out. A tool not on the list isn't counted, and neither is a use the patterns didn't find (an unlisted alias, a link without the name): we read what the patterns matched, so these counts can miss uses but not invent pairs.
• Excerpts, not whole skills. Readers saw the lines around each match, not the whole workflow, so a condition or a "don't" further away can be missed.
• Our own tools are counted as names only; those matches weren't read.
• No causal claim, and nothing about whether a skill that uses a tool is better or worse for it.

Which outside tools do the skills use?

We counted a skill once per tool, then had the matching excerpts of every one of the 335 pairs read and labelled with the first that applies: a step uses the tool, the skill studies it (answer engines only), the skill only mentions it, the skill says not to use it, or the match means something else. A label counts only when both reading runs gave it.

Search Console is the tool a step uses in the most skills (23); Semrush is named in the most (38), but a step uses it in 16

Skills, of 60, for the 6 tools a step uses in the most skills; a skill counts once per tool. All 15 tools are in the table below. One consultancy's 60-skill library, measured 10 Nov 2026.

A step uses it (both reading runs)

1. Search Console: 23
2. Semrush: 16
3. Screaming Frog: 12
4. Webflow: 6
5. WordPress: 6
6. Google Sheets: 6

The tool's pattern matched (named in the skill)

1. Search Console: 36
2. Semrush: 38
3. Screaming Frog: 23
4. Webflow: 20
5. WordPress: 19
6. Google Sheets: 10

• "Matched" counts skills whose files the tool's pattern found. It includes the 2 pairs where the word meant something else (one each for ChatGPT and Google AI Overviews).
• "A step uses it" covers a step that reads data from the tool, does something in it, or puts a deliverable there. "With no alternative or condition" is the part of those where both runs gave the qualifier "required": the step names this tool, offers no named alternative, and isn't conditional ("if you have …"). It's what the text says, not that the job needs this tool: 64 of the 96. Where both runs said "uses it", they gave the same qualifier in 91 of 96.
• "Only mentions it" means the name appears but no step uses the tool, as in background, a comparison, an example or a test prompt. We didn't count which of these each mention is.
• "Runs disagreed" is a pair the two runs labelled differently; it gets no label of its own. Six of the 9 are answer engines, where one run said "studies it" and the other "only mentions it". "Says not to use it" ranks below "only mentions it", so a skill that names the tool elsewhere too counts as a mention: the 1 is a floor.
• Counting skills: 39 of the 60 skills have a step that uses at least one of the 15 tools; 29 use Search Console, Semrush or both.
• What the readers saw: each pair's matching lines with two lines of context, up to 12 places, and long lines cut to 240 characters. Those limits touched 199 of the 335 pairs, so a third, blind read went over all 199 with every place and no cut lines. It kept 188 labels. It moved 4 to "a step uses it" (Webflow 2, Semrush 1, Google AI Overviews 1) and 5 from "uses it" to "only mentions it" (one each for Search Console, Semrush, Ahrefs, Google Sheets and Google Docs); the other 2 moved between labels other than "uses it". With its labels, Search Console would be at 22, Semrush 16 and Screaming Frog 12: the same top three. The counts in the table are the two runs'.

Which of our MCP server's tools do the skills name?

This section is exploratory: these counts are exact-name text matches, not read and not calls, so "names" means only that the name is somewhere in a skill's files. The skills are written for an MCP server with 90 tools: ours, on the date measured. That's our catalog, not an industry list. We grouped the tools by job, by hand (the table is in the method), and counted a skill once per group.

Our MCP server's tools, by job: 23 of 60 skills name none outside their own learning loop

Skills, of 60, that name at least one tool of the group (exact-name text matches, not read); a skill counts once per group. One consultancy's 60-skill library, measured 10 Nov 2026.

1. Scraping and crawling pages (7 tools): 26
2. Keyword and trend research (3 tools): 21
3. Search results (1 tool): 17
4. AI answers and citations (12 tools): 15
5. Projects, files and account (12 tools): 9
6. The skill's own learning loop (6 tools): 57

• The learning loop is in almost every skill: 57 of 60 name at least one of its six tools (reading the skill and its files, reading, adding and compacting recorded learnings, updating the skill).
• Beyond the loop: counting each skill's distinct tools outside the loop, zeros included, the median is 1.5, and 41 of those 84 tools are named by at least one skill.
• The catalog changed while the skills were written. 17 skills contain the name of a tool the server has since retired, 13 of them a keyword planner. These are text matches we didn't read, so some may sit in code or notes rather than in steps.

What it means

In this library, the outside tool a step uses in the most skills is Search Console, which (in our reading, not a measurement) is where a client's own search data lives; and a list of tools found by searching the skill files overstates what the steps use: 203 of the 335 skill-and-tool pairs were mentions. The rule in our published outline, written before the reading, to say more only where most skills that name a tool use it, holds for Search Console (23 of 36) and not for Semrush (16 of 38).

Context, not a finding: where a step reads Search Console data through Google's API, it needs access to the site's property. Google's API documentation says "All requests to the Google Search Console API must be authorized by an authenticated user", and its prerequisites say "Your account must have the appropriate Search Console permission on a given property in order to call that method". We didn't measure how these skills get the data: through the API, an export someone hands over, or another route. In Claude, one way to grant that kind of access is a connector: "Connectors let Claude access your apps and services, retrieve your data, and take actions within connected services."

Methodology

• Data: the 60 skills in one account (the consultancy's own): 55 personal skills and 5 it keeps in client projects, all written for its own client work. They were copied with client, partner and person names replaced by fictional stand-ins. Quoted text in this article is from the cited sources, never from the skills. We publish counts only.
• Matching: one case-insensitive pattern per tool, with aliases ("GSC" counts as Search Console, "GA4" as Google Analytics), searched in every file of a skill except its recorded learnings, with code blocks and URLs included. A skill counts once per tool. A skill whose own name contains a matching word (a "gsc-" prefix) counts as naming the tool; the reading then decides what that name is.
• Dates: "measured on 10 November" is when the scripts ran; "last modified on 5 November" is the newest file's modification time in the copy (UTC). The copy's fingerprint (a SHA-256 over every file's hash and path) begins d89f4acb86ce7d3a, and the tool catalog's acee383795bd043b.
• The reading: every pair's matching excerpts were read, none sampled. A reader saw the skill's description and each matching line with two lines either side (long lines cut to 240 characters, around the match) and the nearest heading above, SKILL.md first, up to 12 places per pair. One label per pair, the first that applies, in the order above. With "uses it", a qualifier: no alternative or condition, this or a named alternative, or only if available. The readers were Claude Opus 5.5 sessions, each fresh, given only a written protocol and its batch.
• Two runs: a second set of sessions of the same model labelled every pair blind. The runs agreed on 326 of 335 (Cohen's κ = 0.949 over the five labels). That measures consistency between two runs of one model, not correctness. A pair has a label only when both runs gave it; the 9 others get no label (the table's "Runs disagreed"). No person checked the labels.
• A third read, for the limits: 41 pairs had more than 12 matching places and 192 had a cut line (199 had either way: an excerpt that differs from the full one). Fresh sessions of the same model re-read all 199, blind, with every place and no cut: hidden text can hide a use, and hidden context can also undo one. They kept 188 labels; 4 moved to "uses it", 5 from "uses it" to "only mentions it", 1 from "only mentions it" to "studies it" and 1 to "says not to use it". The readers of the 66 pairs first labelled "uses it" got one extra sentence: an example, a quoted or retired workflow or a conditional mention isn't a use. That sentence sits uneasily with the "only if available" qualifier, and only those readers got it, so the check isn't uniform. It's one run, a sensitivity check; the published counts stay the two runs', with the check's beside the headline ones.
• Left out: one partner's AI platform, whose name the scrub replaced. The pattern finds it in 25 skills; we left it out of the reading and the charts so it can't be identified. Its use wasn't measured, so "the most skills" is among the 15 tools we read.
• Our tools: a skill names a tool when the tool's exact name appears, alone or after a prefix such as mcp__…__. These counts weren't read. The grouping by job: scraping and crawling pages (7 tools), keyword and trend research (3), search results (1), technical and on-page audits (2), backlinks (1), web research (1), AI answers and citations (12), competitor monitoring (5), ads, audiences and social (6), client context and knowledge base (20), projects, files and account (12), skill library (14), and the learning loop (6). Every tool is in exactly one group.
• Reproducibility: three scripts count every number from the copy and the private labels (the library's size, the account's skill split, and these gates) and write counts only. They record a fingerprint of the copy, of each run's labels and of the tool catalog, and the result re-cuts the reading packets to check they match the labels. The copy and the labels stay private, so the numbers can be audited by us but not re-run by a reader. Sources read on 10 November 2026; the quoted sentences are stored with the result.

Disclosure

Unilytics builds a skill system for consultants and the MCP server measured here. This library is ours and lives in it. We chose the 15 tools, wrote the labels and the protocol, and left out the partner platform. The labels were made by Claude through Anthropic's API. Corrections: if a number here is wrong, we fix it on this page with a dated note at the top.