Another SEO meeting, a lot of words, and still no way to say what changed?

Croftweaver for established sites: open-source skills that give your team the search and AI work on a live site. The checks and measurements work on any stack; the ready-made test suite needs the Croftweaver starter. They check where the site stands, fix the pages Google already shows, and measure what changed.

Know what to fix. Check what changed. Fast.

Reports arrive. Changes wait.

On a large site, search work often runs through a retainer, a ticket queue and a monthly report. A title change waits for a sprint. A report says what moved, but you cannot rerun it, and you cannot see what the hours bought.

This is not about any one agency. It is about a way of working that makes the site slow to change and hard to inspect, and that leaves the people who own the result unable to check it.

It is your site. You should be able to see where it stands, change what the data points to, and check the result yourself. Agencies can use these skills too.

What stays wrong if nothing changes

A method your team can inspect

Croftweaver is a set of open-source skills for AI coding assistants such as Claude Code and Codex. The steps read the public pages, the site’s own Search Console data and the answers of AI models, and each ends in a result that can be checked again. The code, the saved data and the keys sit with your team. The licence is MIT.

The evidence so far comes from one small site, with its limits stated on the proof page. Nothing here promises a ranking.

Get in the driver seat

Sit in the middle of the feedback loop and drive it, week after week. Check, fix, measure, and around again. Your assistant keeps the loop turning. You decide what changes and you see the result yourself.

Not a report from an outside company that is hard to follow and slow to arrive. The loop runs on your own search data, in your own repository, at your own pace.

A loop of three steps, check, fix and measure, with you in the middle. Check Fix Measure You

Three steps

  1. 1. Check

    See where each target search ranks and what four AI assistants say about the business. Read-only, with no change to the site.

  2. 2. Fix

    Take the 20 to 30 pages Google already shows. Write guard tests first, change the pages, and optionally have other AI models review the change.

  3. 3. Measure

    Three to six weeks later, set the prediction next to the outcome, and keep the weekly check running.

The same plan in six steps, in this order

On a large site, the best place to start is the pages Google already shows. The order makes sure those are found before anything new is written.

  1. Baseline: where the site stands today

    search-console-insights

    • The position of each target search, from Google Search Console, after a one-time, read-only sign-in.
    • Searches just below page 1 (positions 8 to 20). These are candidates worth investigating first, not a promise of an easy gain.
    • Pages shown often but rarely clicked. The skill checks the live results page before it suggests a new title or description.
    • Google and Bing side by side. A weekly tracker builds the history, and a report page shows clicks from Google per week for up to 16 months.

    Say to the assistant: Connect Search Console for example.com and show me the searches just below page 1.

  2. AI visibility: does an AI assistant name the business?

    search-console-insights: geo_check.py

    • Asks ChatGPT, Claude, Gemini and Perplexity, and optionally Google AI Mode and AI Overview, your customer’s own questions, without naming the business.
    • Where an engine allows it, two answers: “knows you” from memory, and “finds you” with web search on, which lists the sources the engine returned. Gemini gives the first only. Perplexity through OpenRouter and Google’s own AI answers give the second only. For Perplexity, the AI check reports how often your site was among the sources it returned, not how often it was cited, because the check does not tell which of those sources the answer quotes.
    • Three question slots: broad, narrow and branded. The branded answer is read for accuracy and is not scored.
    • Needs no Search Console access. It needs a key for each assistant you ask, or one OpenRouter key for all four. On a large site, use each assistant’s own key: your market’s country is then sent with the searches of ChatGPT, Claude and Perplexity. The toolkit sends no country through OpenRouter today.

    Say to the assistant: Set up the weekly AI check for example.com with these two customer questions and one question about the company.

  3. Diagnose

    seo-auditwebsite-positioning-checkseo-reposition

    • A technical audit: crawling, indexing, canonical tags, language and country tags (hreflang), Core Web Vitals.
    • A positioning check with a one-screen verdict of Clear, Blurred or Broken, the biggest blur, and the three fixes that would change the most.
    • When searches without the brand name bring almost nothing, the trap test reads each candidate phrase against the live top 10 and marks it wedge, crowded or trap. A person confirms every call.

    Say to the assistant: Run an SEO audit on example.com, starting with the service pages.

  4. Fix the pages that are already seen

    schema-markupai-seocopywritingsite-architecture

    • Start with the 20 to 30 pages that already have impressions, before any new page.
    • Structured data such as Organization, Service and Product as JSON-LD, with notes for sites on a content system.
    • Facts an AI assistant can quote: real figures in text and tables, each with its source.
    • Hubs and internal links, so deep pages are a few clicks from the home page.

    Say to the assistant: Add Organization and Service structured data to the service pages.

  5. Guard and review

    seo-repositionindependent-reviewdouble-knuth

    • Guard tests come before the copy changes. They start red, turn green page by page, and stay as regression checks. Each guard is first shown to fire on a broken sample.
    • Optionally, plans and diffs go to independent reviewers from other model families before anything is built or merged. You choose whether to use an outside reviewer.
    • A second pass reads the finished change as a whole for drift across files, which a review of the diff alone does not cover.

    Say to the assistant: Write the guard tests from the wording guide before we touch the copy.

  6. Measure again

    seo-repositionsearch-console-insights: track.sh

    • A grading run is booked 3 to 6 weeks after shipping. It pulls Search Console and the live results again and sets the prediction next to the outcome.
    • The weekly AI check shows the trend per engine and marks a week as not comparable when the question, the model, the settings or the route changed. Note any other change made in the same period: the grading run cannot separate them.

    Say to the assistant: Schedule the grading run for six weeks from today.

What your team needs

ItemStepsNote
Search Console access to the site1, 3, 6One-time setup: a Google Cloud project and a read-only sign-in. You may need to sign in again later. The token is stored on the computer that runs the skills.
OpenAI, Anthropic and Perplexity keys2Recommended for a large site. The country you set is sent with each search. If an OpenRouter key is also set, it takes over all four assistants unless you name the ones that should use their own key.
Google Gemini key2Asked from memory only, never with web search. Google’s terms may require a paid account in the European Economic Area, the UK and Switzerland.
OpenRouter key2One key for all four assistants: the simplest start. The toolkit sends no country through it today. OpenRouter documents a country setting for web search, but it is in beta, works only with a provider’s own search (it is ignored with OpenRouter’s Exa, Firecrawl, Parallel and Perplexity search), and the toolkit does not use it. Checked in OpenRouter’s documentation on 9 October 2026.
Serper key3, 6Optional, free tier. The live top 10 for any search.
SerpApi key2, 3, 6Optional. Adds Google AI Mode and AI Overview, with your country and language sent with the search. It can also serve the live top 10 in place of Serper.
Bing Webmaster key1Optional, free. Adds the Bing column.
Write access to the code4, 5Only when changes are made.

It runs with Python 3.9 or newer and an AI coding assistant. The weekly schedule helper is for a Mac. On Linux the tracker runs from cron or a systemd timer that you set up yourself. Windows is not covered.

Where your data goes

Check each provider’s terms and retention policy against your company’s rules before you send confidential questions.

Start with a four-week pilot

Do not start with the whole site. Pick one section, the 20 to 30 pages with the most impressions, and one person who approves changes. Croftweaver does not replace your change process: run the edits through your own staging, approval and rollback steps, and get your security team to approve the AI coding assistant first. This is a plan for where the time goes, not a promise of a result.

  1. Week 1

    Connect Search Console and pull 90 days. Write two questions your customers ask and one about your company, and run the first AI check.

  2. Week 2

    Audit the 20 to 30 pages with the most impressions. Run the positioning check on the home page and the main offer pages.

  3. Week 3

    Write the guard tests, then fix titles, structured data and quotable facts on those pages.

  4. Week 4

    Review, ship, and book the grading run. The weekly AI check keeps running.

After the grading run

Weave your website with AI

If your company wants to become AI-driven, the website belongs in that plan. AI assistants answer questions about companies, and some answers cite websites. Croftweaver helps you prepare your pages for those answers and check whether your business is named or cited. And your own team can use an AI assistant to change the site, instead of waiting for the next ticket.

That is the ambition behind Croftweaver: weave the website with AI in both directions. Let AI read it, and let AI help change it.

A good first AI project, alone or with your team

AI stops being abstract when it works on your own site. You know what it should say, you see each change take shape, and you decide what stays. In my experience it is a real learning curve, and a happy one. Questions along the way? I am glad to help. Write to .

What it does not do yet

The roadmap says what comes next and what stays out on purpose.

Things change fast. The AI providers change their tools, terms and prices often, and so does Croftweaver. We updated this page on and will review it regularly. Please do your own research: check current provider terms, pricing and country support before you set anything up.

Questions and answers

Getting started

Is SEO still worth it with AI?

Our view: yes, as the base. The AI check lists the pages that assistants with web search return, so a site that is correct, well structured and easy to find should serve both search and AI answers. What changes is how you are found: by link in a list, and by name in an answer. The AI check shows the second one for your own company.

Where should I start?

With the pages Google already shows. Connect Search Console, pull the last 90 days and list the pages with the most impressions. Write two questions your customers ask and one about your company, and run the first AI check. Then work on the top 10 pages before you plan new ones.

How do we pick the first 10, 25 or 50 pages?

Your assistant sorts the Search Console data: pages with many impressions, searches at positions 8 to 20, and pages that are seen often but rarely clicked. Your team adds what the data cannot know: which pages sell, and which are about to change. Start with 10, and widen to 25 and then 50 once the first changes are measured.

What do we need before we start?

Three things: access to the site’s Search Console, which needs a one-time setup with a Google Cloud project and a read-only sign-in; an AI coding assistant; and keys for the AI services the weekly check asks. On a large site, use each assistant’s own key, because only those send your market’s country with the search. One OpenRouter key is the simpler start. Optional keys add Bing, the live top 10 and Google’s own AI answers. Write access to the code is needed only when changes are made.

Do we need a new site or CMS first?

No. The checks and measurements work from the live pages and your Search Console data, on any system. You keep your CMS. The ready-made test suite needs a site built with the Croftweaver starter.

How long until we see results?

We do not promise a result or a date. Plan on about four weeks for a first pass on one section: a baseline, a first AI check, fixes on the 20 to 30 pages already seen, and a review. A grading run 3 to 6 weeks after shipping sets the prediction next to what happened. Search moves slowly, so read it over weeks.

Who does the work

Can AI replace my SEO agency?

In our view AI can take over much of the repeatable agency work, but not all of it, and Croftweaver does not try to replace the rest. AI is useful for the repeatable work: audits, structured data, drafts, checks and reports. On a corporate site it can do much of that work with people in charge. People still set the direction, decide which markets are worth pursuing, approve what the brand and legal team can say, build relationships and answer for the result. What changes is what you can check yourself: where the site stands, and what moved.

Can AI do SEO for a corporate site?

Croftweaver is built so that your assistant can do much of the work, with people in charge. It gives the assistant these steps: check the live site, find the pages Google already shows, add structured data, write guard tests and measure the result. It cannot decide which markets are worth pursuing, check a claim against your legal rules or earn trust for your brand. Those stay with your team and your partners.

Which work stays in-house?

Run the two measuring loops in-house: Search Console and the weekly AI check. The data is yours, and the numbers should not depend on who is paid to improve them. Keep what only you can know or approve in-house too: product facts, brand voice, legal sign-off, access to your experts, priorities and the technical platform. An agency adds most when it reads the results with you, benchmarks you against competitors, produces content at volume, and handles outreach and public relations. Both sides then work from the same numbers.

How should a company and its agency work together?

Share the facts and the checks, and change the site in small steps. The company hands over approved facts and brand rules once. The agency, or your own assistant, proposes each change as a draft your team can review. Guard tests, where the setup allows them, are written before the copy changes and say what done means. The weekly check shows what moved, and your own staging and sign-off decide what goes live.

How do new answers and page updates reach our CMS?

The usual route is the CMS’s own interface for programs (its API), with an account that can write drafts but cannot publish; confirm that your CMS can do that before you rely on it. A workable setup is that the assistant writes each answer as its own entry, with its question, its answer, its source, its owner and the pages it belongs on. Your team reviews the draft and publishes it as usual. Croftweaver ships no connector for every CMS; connecting one is setup work.

How does an assistant work through a site with hundreds of pages?

A setup that works: it reads the content from two places and keeps a copy of both in one repository that a weekly job refreshes. Croftweaver does not ship that job yet (see the last question); you build it for your site. The CMS stays the source of truth: every change goes back into it as a draft, never into the copy. Scripts do the sweeps across all pages; the assistant reads the results, then works page by page, starting with the pages Google already shows.

MethodWhat you getVerdict
Your CMS’s API (WordPress, Contentful, Storyblok, Sanity and others)What editors wrote, usually as separate fields: title, body, Q&A entries, language versions, entry IDs, last-changed dates, unpublished draftsThe main source. It is also the channel the assistant later writes drafts back through, where your CMS allows it.
The sitemap and the published pages (what the Croftweaver facts check reads)What Google and AI assistants actually see: page templates, meta descriptions, structured data, footers, cookie textAlways add it. Set next to the CMS, it shows problems that sit in the templates, not in the content. It needs no access, so it also works on a competitor’s site.
A full offline copy (SiteSucker, wget, HTTrack)Every file, images and scripts included, but no IDs, no dates and no structureOnly for a one-off look. A frozen copy is the wrong base for a weekly loop, and much of what it stores is noise.
A Search Console exportWhich pages matter: impressions, clicks and position for each pageIt joins onto both, and it decides where the work starts.

If a site builds its text in the browser rather than on the server, a plain fetch sees empty pages, and a headless browser such as Playwright is needed. Many sites, including many built with Next.js, send their text from the server, so a plain fetch works; compare a plain fetch with the rendered page for your site before you rely on it.

Visibility in AI answers

What is SEO for AI called, and what does it look like for a large company?

It goes by several names: GEO (generative engine optimization), AEO (answer engine optimization) or AI SEO. It builds on SEO, not in place of it. For a large company it is a loop: a baseline of where the site ranks, a check of what assistants say about the company, fixes on the pages already seen, and a new measurement weeks later. It tends to work better when every page and every profile states the same facts.

What hurts our visibility in AI answers?

Three things come up most often.

Answers that are hard to quote. Give an assistant short passages that stand on their own: a clear answer at the start of a section, a figure with its source, a table, a question with its answer. Split long marketing paragraphs into passages like these, and put every figure in the text as well as in any image. Structured data tells machines what a page is about, such as the company, a service or a product. It supports the visible text and does not replace it.

Facts that disagree. If one page says 500 clients and another says 700, an assistant may quote either, or neither. Keep one list of approved facts, each with a source and an owner, and have your assistant compare your pages with it. The weekly AI check shows what assistants already say about you, including any wrong figure they repeat.

A blurred position. If your pages describe the company in different words, an assistant may not be able to tell what kind of company you are or whom you serve. Croftweaver works out your positioning before any copy is written, and a check lists the pages in your rules that have lost it. For sites built with the Croftweaver starter that check is a test that runs on every change; for other sites the check reads the published pages.

How does that work on a corporate site with several audiences?

A large company often speaks to several groups at once, for example customers, job applicants and investors. Each group gets its own positioning: what it would use instead, what you offer it, and why it should believe you. What stays the same for every group is the company itself: one name, one category, one set of facts.

Every page then belongs to one audience and carries that audience’s term in its title, its description and its main heading or opening paragraph. Where your CMS allows it, keep the audience as one field per page. A check reads the published pages in your sitemap (up to 1,000 by default), on staging or on the live site, and lists the pages your rules cover that have lost their term. Pages you exempt, such as legal pages, pages past the limit and pages it cannot read are not checked. Your team decides whether the page or the term needs to change.

The weekly AI check asks two questions per site today, plus one about your company. Your assistant drafts them from your homepage and, if you have one, your positioning, and you confirm or change the wording. If you are unsure which audience to start with, ask it to suggest one from your pages and the groups you serve. After you switch to another audience, the new weeks cannot be compared with the old ones, and checking several audiences side by side is not built yet.

Croftweaver ships the positioning method, and the test for sites built with its starter. The check of the published pages on any other system is part of the facts check.

How can AI keep our facts consistent?

Your team keeps one list of approved facts, each with a source and an owner. The facts check reads the pages in your sitemap, or a list of pages you give it, up to 1,000 by default. For each fact you list the words that go with it, such as “clients”, and the check compares the numbers it finds next to those words with the approved value. It sets aside a 0 unless the approved value is 0, and a plain year from 1900 to 2100 right after a © or a word such as “in” or “since” (“founded in 2016”) unless the approved value is in that range. It reports every mismatch with the page and the sentence it is in. It lists the pages it could not check, such as pages that returned an error, pages blocked by robots.txt and PDFs. It also looks for old names and claims that are still in use, and does the same for the profiles on review sites and directories that you add. The weekly AI check’s question about your company, read for accuracy, shows which wrong figure an assistant already repeats. After the cleanup, run the check weekly, and again after any change to a fact.

How can we tell whether AI assistants name our company?

Every week, ask the same questions your customers ask, without naming the company, and count how often each assistant names it. The AI check does this for ChatGPT, Claude, Gemini and Perplexity, and optionally Google’s AI Mode and AI Overview. Where an assistant allows web search, the check also lists the sources it returned; Gemini is asked from memory only. A third question asks what each assistant knows about your company; it is read for accuracy, not counted. The check counts mentions. It does not tell you whether an assistant recommends you.

What if assistants quote directories and review sites instead of our own pages?

Then those profiles are the next job. With web search on, the AI check lists the sources each assistant returned, so you can see which sites they are (Gemini is asked from memory only; the check does not request web search from it). If the sources are directories, review sites or press pages, make sure your profile there states the same facts as your site and links to it. Croftweaver covers directory and profile links, and requests for links from existing partners, sponsors and speakers. It does not do cold outreach or bought links.

Safety, cost and scope

Is it safe to let an AI assistant change a corporate site?

Treat it like any other contributor. The assistant proposes a change, your team reviews it, and it goes through your own staging, approval and rollback. Have your security team approve the assistant first, and check each provider’s terms before confidential questions leave your network. The scripts make no call to a Croftweaver server, and the Search Console sign-in stays on your own computer.

Who inside the company needs to be involved?

Four roles matter. Your security team approves the AI assistant and checks the providers’ terms. The owner of the site or CMS gives access and runs staging and rollback. Brand and legal approve what the pages may say. And one person approves each change. In a four-week pilot on one section of the site, that can be a small group.

Which data leaves our network?

The questions of the AI check, two that your customers ask and one about your company, go to the AI services you choose: straight to OpenAI, Anthropic, Google and Perplexity with their own keys, or through OpenRouter. If you use SerpApi for Google’s AI answers, those questions go there too, and the search phrases for the live top 10 go to Serper or SerpApi if you use one. If you use the optional outside review of plans and changes, the text under review goes to the reviewing services too. Your assistant sees what you give it. The Search Console sign-in, the history and the saved answers stay on the computer that runs the skills. Nothing goes to Croftweaver. Check each provider’s terms before confidential questions leave your network.

Does Croftweaver replace our SEO tools or our agency’s reports?

No, it sits beside them. It adds checks you can run yourself on the public pages and your own Search Console data, so any report, from anyone, can be set next to numbers you can rerun.

What does it cost?

Croftweaver is open source and free to use. You pay for your AI assistant and for the AI services the weekly check asks. Search Console and Bing Webmaster Tools are free.

What can Croftweaver already do today?

Ready today, on any live site:

  • Search Console and Bing reports: where each target search ranks, searches just below page 1, pages that are seen often but rarely clicked, and a weekly history.
  • The weekly AI check: two questions your customers ask and one question about your company, asked of ChatGPT, Claude, Gemini and Perplexity, and optionally Google’s AI Mode and AI Overview.
  • The facts check: the numbers that sit next to the words you list for each approved fact, compared with the approved value on the pages of your sitemap (up to 1,000 by default), plus old names and claims that are still in use.
  • The positioning check of live pages: the pages in your sitemap that your rules cover (up to 1,000 by default), each against the term its audience rule says it owns, in the title, the description and the heading or opening paragraph. Pages you exempt, pages past the limit and pages it cannot read are not checked.
  • Skills your assistant follows for a technical SEO audit, structured data, answers AI can quote, a positioning check and copy.
  • Reviews of plans and changes by independent AI reviewers before anything is built or merged.

Ready for sites built with the Croftweaver starter:

  • The full test gate, including the positioning test and the tone test, on every change.

A method you apply, written for your site:

  • Guard tests on other systems.
  • A separate positioning for each audience: the method is run once per audience.
  • The connection to your CMS through its API.

Not built yet:

  • A weekly snapshot of every page’s text, so you see what changed on the site.
  • More than two customer questions per site in the AI check.
  • A ready-made list of the first 10, 25 or 50 pages: today your assistant builds it from the Search Console report.
  • Moving an existing site onto the Croftweaver starter.

The example prompts use example.com. Replace it with your own domain.