The AI
Head of SEO

claude mcp add --transport http chatseo https://api.chatseo.app/mcp

FUNCTION 01 — Search Console analysis
Feeds: Performance report · Weekly action list

Pull the last 28 days and the previous 28 from Search Console for {site}: clicks, impressions, CTR, position, by page and by query. Exclude brand queries first and say how many rows that removed.

1. State the site's totals for both periods and the delta, brand excluded.
2. Cut the query set into position bands: 1-3, 3.1-10, 10.1-20, 20+. For each band: number of queries, clicks, band CTR.
3. Name the ONE band where the gap between impressions and clicks is largest in absolute clicks. That is the constraint.
4. Inside that band, list the 5 queries above 100 impressions with the lowest CTR relative to the band average, with the page each one lands on.

Output: the constraint in one sentence with the numbers behind it, then the 5 queries, then ONE action on ONE page — the change, the metric that will move, the date to check.
Do not list alternatives. If a figure is not in the data, write "cannot compute from this data".

FUNCTION 02 — Keyword research
Feeds: Keyword priorities

Start from what Google already shows us for, not from a tool's volume column. Pull every query with 100+ impressions over the last 90 days, brand excluded.

1. Cluster queries that share one intent and could realistically be answered by one URL. 3+ queries = a cluster, named after its highest-impression query. Fewer = merge into a neighbour or drop.
2. For each cluster: total impressions, best current position, the URL of ours that currently ranks (or "none").
3. Classify intent: informational / comparison / transactional / navigational. Drop navigational.
4. Rank clusters by: impressions × (1 if we have a URL at position 4-20, 0.5 if we rank 21+, 0.3 if no URL). Show the arithmetic.

Output: the top 10 clusters as a table (cluster · intent · impressions · our URL · position · score), then ONE cluster to work this month and the one reason it beat #2.
Do not invent search volumes. If a number is not in the data, write "cannot compute from this data".

FUNCTION 03 — Competitor analysis
Feeds: The SEO strategy · Keyword priorities
Needs: 3 competitor domains from me. Ask if missing, then stop.

Take the top 20 non-brand queries by impressions where we rank 4-20.
1. Fetch the live top 10 for each if you have a search or fetch tool. If you do not, ask me to paste the top 10 URLs per query and stop.
2. Record which of the 3 competitors appear, at what position, and the page type they used: guide / comparison / product / tool / other.
3. Count, per competitor: on how many of the 20 queries do they outrank us, and with which page type most often.

Output: one table (query · our position · competitor · their position · page type), then the single pattern — the ONE page type or topic where one competitor beats us most consistently — and ONE page we should build or rewrite to take it, with the query it targets.
Do not estimate domain authority, traffic or budgets. If a fact is not visible in the SERP or in our data, write "unverifiable".

FUNCTION 04 — Technical audit
Feeds: Fix queue

Audit only what blocks clicks on pages that already earn impressions. Pull the top 50 pages by impressions, last 28 days.

For each page, using fetch or crawl tools if you have them (if not, ask me for a crawl export and stop), check:
- HTTP status and any redirect chain
- indexable: no noindex, canonical points to itself
- title and H1 present, and unique across the 50
- response time or LCP, if measurable
- mobile rendering problems visible in the HTML

Rate every finding: BLOCKING (the page cannot rank or be indexed) / DEGRADING (ranks but loses clicks) / COSMETIC.

Output: counts per rating, the BLOCKING findings listed with URL and the exact fix, then ONE fix to ship first — the one touching the most impressions. Effort in hours.
Ignore COSMETIC entirely. Do not report an issue you did not observe on a fetched page. If you could not fetch a page, write "not checked".

FUNCTION 05 — On-page optimization
Feeds: Fix queue · Changes executed
Needs: one URL from me. If none is given, take the page named by the latest Search Console analysis.

1. Pull this page's queries, last 28 days: impressions, clicks, CTR, position. Brand excluded.
2. Fetch the live page. Record the current title, meta description, H1, H2s and word count.
3. Identify the ONE query with the most impressions where the page sits at position 3-15 and the title does not contain the query's core terms.
4. Rewrite the title (60 characters or fewer) and the meta description (155 or fewer) around that query. Keep the brand suffix if there is one. Propose an H2 change only if a query above 100 impressions has no matching section on the page.

Output: current vs proposed title and description side by side, the query and its numbers, the metric that should move (CTR in that position band) and the date to check it (28 days). Then stop.
One page, one query, one rewrite. Do not touch body copy unless asked. Do not promise a ranking change.

FUNCTION 06 — Content planning
Feeds: Content plan
Reads: the clusters from Keyword research (02) and the winning page type from Competitor analysis (03). If 02 has not been run this month, run it first.

1. Take every cluster where we have no URL, or our URL ranks 21+.
2. For each: intent class, the page type that wins its SERP (from 03, or "unknown" if 03 has not run), the cluster's total impressions, and whether an existing page could be extended instead of a new one built.
3. Score: impressions ÷ effort in hours (new page = 8h, extension = 3h). Show the numbers.
4. Cap the plan at 4 pieces per month.

Output: this month's 4 pieces as a table (working title · target cluster · page type · new or extend · effort · impressions at stake), each with a 2-line brief: the question the page answers, and the one thing it must contain that the competitors' pages lack. Then name the ONE to write first.
Do not add a piece without a cluster behind it. Do not invent impressions for topics we have never appeared for.

FUNCTION 07 — Internal linking
Feeds: Fix queue

1. Pull pages by impressions, last 28 days. Split them: STRONG = the top 10 pages by clicks. TARGET = pages ranking 4-15 on a query with 500+ impressions.
2. Fetch each STRONG page and list its outgoing internal links. If you cannot fetch, ask me for a crawl export and stop.
3. For each TARGET page: which STRONG pages are topically related (they share a query stem or a section topic) but do not link to it.
4. For each missing link, propose the anchor text — the target's main query in natural wording — and the exact sentence in the STRONG page where it fits.

Output: a table of at most 10 links to add (from · to · anchor · where in the page), ordered by the TARGET page's impressions. Then ONE link to add today.
Do not propose links from pages you did not fetch. Do not propose navigation or footer changes.

FUNCTION 08 — Schema markup
Feeds: Fix queue · Changes executed
Needs: one URL, or take the page from the latest On-page run.

1. Fetch the page. Extract any existing JSON-LD and list its @type values.
2. Decide the ONE schema type the page's content actually supports: Article, Product, FAQPage (only if the page already shows visible questions and answers), HowTo (only with visible steps), Organization, or none.
3. Write the complete JSON-LD block using only values visible on the page: headline, dates, author, prices, questions. Do not invent ratings, review counts or dates.
4. State where it goes (inside the head) and how to verify it: Google's Rich Results Test, paste the URL.

Output: the JSON-LD in a code block, the type you chose and one line on why the others do not apply, then the one thing to check in 28 days: this page's impressions under Search Console's search-appearance filter.
If the page supports no type honestly, write "no schema — nothing on this page qualifies" and stop.

FUNCTION 09 — Traffic drop diagnosis
Feeds: Weekly action list · Performance report
Run when clicks are down 15%+ week on week, or when I ask.

1. Confirm the drop: last 7 days vs the previous 7, AND vs the same 7 days last year. If last year shows the same dip, label it SEASONAL and stop.
2. Find where it lives: by page, by query, by country, by device. Name the single dimension that explains the largest share of lost clicks, with the number.
3. Inside that dimension, check which applies: impressions fell too (demand or indexing) / impressions held and CTR fell (SERP layout or title) / impressions held and position fell (ranking). Check that the affected pages return 200 and are indexable.
4. Label the cause: CONFIRMED / PROBABLE / UNVERIFIABLE, with the figure or the test that would confirm it.

Output: the drop in one line with both comparisons, where it lives, the labelled cause, and ONE action — or "hold, recheck on {date}" if the cause is UNVERIFIABLE.
Do not attribute a drop to an algorithm update without a dated source. Do not estimate lost revenue.

FUNCTION 10 — Opportunity detection
Feeds: Weekly action list · Keyword priorities

Find the cheapest clicks on the site right now. Pull queries, last 28 days, brand excluded.

Score three signals and show the arithmetic:
A. Striking distance — queries at position 4-10 with 500+ impressions. Value = impressions × (this site's position 1-3 CTR − the query's current CTR).
B. Page-two intent — queries at 11-20 with 1,000+ impressions where we have a page that mentions the query. Value = impressions × this site's position 6-10 CTR.
C. CTR gap — queries at 1-3 whose CTR is 30%+ below this site's own 1-3 average. Value = impressions × (average − actual).

Output: the top 5 across A, B and C as one table (query · page · position · impressions · signal · value in clicks), then ONE opportunity to take this week and the exact change that takes it.
Use this site's own band CTRs, never a published CTR curve. If a band has fewer than 30 clicks, write "below the floor" for it.

FUNCTION 11 — AI search visibility
Feeds: The SEO strategy · Performance report
Needs: the 10 questions a buyer asks before choosing a product like ours. Ask me if missing, then stop.

1. For each question, query the AI surfaces you can reach (a search tool that returns AI answers; otherwise ask me to paste the answers from ChatGPT, Perplexity and Google's AI Overview). Record: are we cited, is a competitor cited, which URL, and what claim the citation supports.
2. For every question where a competitor is cited and we are not, fetch their cited page and name the specific thing it has that ours lacks: a direct answer in the first 100 words, a comparison table, named numbers, a dated update, a named author.
3. Count: cited / not cited / not checked.

Output: a table (question · cited: us / competitor / none · URL · what the cited page has that ours lacks), then ONE page to change and the ONE element to add to it.
Do not claim how any AI system ranks or selects sources. Report only what the answers actually showed. Anything you could not check is "not checked".

FUNCTION 12 — Fix execution
Feeds: Changes executed
Takes: the top item of the Fix queue, or the fix I name.

1. Restate the fix in one line: page, change, the metric expected to move, the check date.
2. Produce the change in the exact form it ships: the new title tag and meta description as HTML, the JSON-LD block, the internal link with its anchor and the sentence it sits in, or the redirect rule. Nothing left to paraphrase.
3. If you have a publishing tool connected to this site's CMS: apply the change, then fetch the live page and confirm the new value is present. If you do not: output the change and the precise place to paste it, and stop.
4. Log it as one row I can paste into the Changes executed list: date · page · change · expected metric · check date.

Output: the ship-ready change, the confirmation (or the paste instruction), the log row. Then stop.
One fix per run. Do not batch. Do not touch anything that was not in the fix.

FUNCTION 13 — Scheduled SEO tasks
Feeds: every output, on time

Set up the standing schedule for {site}. For each entry: the function, its cadence, its trigger, and what it hands off to.

Weekly, Monday: 01 Search Console analysis → 14 Performance tracking → 15 Weekly SEO priorities.
Weekly, any weekday: 12 Fix execution on the top item of the Fix queue.
Triggered: 09 Traffic drop diagnosis when clicks fall 15%+ week on week.
Monthly, the 1st: 02 Keyword research → 10 Opportunity detection → 06 Content planning.
Monthly, the 15th: 04 Technical audit → 07 Internal linking.
Quarterly: 03 Competitor analysis → 11 AI search visibility → rewrite the SEO strategy.
28 days after every executed change: re-pull the metric named in its log row and mark it MOVED / FLAT / WORSE.

If you have a scheduling tool, create each of these as a scheduled task carrying the exact function prompt. If you do not, output the schedule as a table I can put in a calendar, and tell me plainly that you cannot run them yourself.
Output: the schedule, then the next task due and its date.

FUNCTION 14 — Performance tracking
Feeds: Performance report
Reads: the Changes executed log. Ask for it if you do not have it.

1. Site level, brand excluded: clicks, impressions, CTR, position — last 28 days vs the previous 28, and vs the same 28 last year. Conversions from analytics if it is connected, otherwise write "not connected".
2. For every change in the log whose check date has passed: pull the metric named in its row, before vs after, and mark it MOVED (10%+ in the intended direction) / FLAT / WORSE. Quote both numbers.
3. Rankings: the 10 target queries from Keyword priorities — position now vs 28 days ago.
4. AI visibility: the cited / not cited count from the last AI search visibility run, if one exists.

Output: four short blocks in that order, every figure with its comparison, then ONE sentence: the single thing this month's numbers say next month is about.
Do not attribute a rise to a change whose dates do not line up — write "timing does not support attribution". No projections.

FUNCTION 15 — Weekly SEO priorities
Feeds: Weekly action list
Run on Monday, after 01 and 14.

1. Gather the candidates: the action from 01, the top opportunity from the latest 10, the top Fix queue item, any open drop from 09, and the next piece from 06.
2. Score each: impressions at stake ÷ effort in hours. Show the numbers.
3. Pick at most 3. Rank them. The first one is the week.

Output, in this exact format, under 200 words:
THIS WEEK: the one action — page, change, who does it, the metric that moves, the check date.
ALSO, IF TIME: two more, one line each.
NOT THIS WEEK: what you dropped and why, one line each.
CHECKING: changes whose 28-day check falls this week, with their metric.

No introduction, no summary, no fourth item. If the data does not support a confident first action, say so and give the cheapest diagnostic instead.