An answer engine reads your page and lifts a paragraph out of it, so the unit that gets cited is smaller than the article. On 2026-10-03 we ran claude-seo's geo command over one published post of ours, and it came back with 72 out of 100 in 87 seconds. The useful half of that report is the passage list: four sentences of ours it called quotable as they stand, and four it said nothing could use. The site around the post scored 82 in a separate run the same day. Neither number says whether anybody cited us, and this lesson keeps that line visible throughout.
The paragraph is the unit, so each one carries its own subject
A quotable passage names its subject, its date and its figure inside itself, because whatever lifts it drops everything around it. A sentence that opens with "this approach" or "the fifth one" arrives at the answer engine with its antecedent missing, and it cannot be used. The test is mechanical: read any paragraph of yours with the rest of the page covered, and ask whether a stranger would know what it is about.
This is the same discipline that makes a page readable, which is why the scores move together. The free readability checker reads a URL and reports sentence length, paragraph length and passive voice without installing anything, and a page that scores badly there usually has passages nothing can lift either.
Our post scored 72, and the breakdown says where
The geo command weights five criteria, and ours landed at citability 17 of 25, structural readability 15 of 20, multi-modal content 8 of 15, authority and brand signals 13 of 20, and technical accessibility 19 of 20. The shape of that distribution is the finding. Our technical base was close to full marks and our formats were the weakest column on the page.
Citability lost points for having no definition and no comparison table, with the first-hand dated facts carrying the rest of the score. Structural readability lost points for having no table while its headings, intro list and visible FAQ passed. Multi-modal lost the most: two annotated screenshots and nothing else, no table, no chart, no video. Authority lost points for the future publish date lesson 11 chased and for having only four outbound source links in the whole article.
The strong passages all carried a date and a name
Four passages were marked quotable as they stand, and they have one thing in common. The opening about the registry gives a date, a server name and a version: on September 27, 2026, the product went live in the official MCP Registry under a specific name at version 1.0.0. The registry section names the two artefacts the process needs. The OAuth section tells a complete small story: a 401, a listing marked unhealthy, and a test account that fixed it. The channels section gives two directory names with the dates attached to each.
Nothing in that list is a writing trick. They score because a reader who arrives at that paragraph alone gets a whole fact, with enough proper nouns that a system can tell what the fact is about. Write the sentence you would want quoted, and put the date in it.
The weak passages pointed at something the reader had to go find
Four passages were called unusable, and each one fails for a reason you can name. The first sixty words of the body narrate a weekend instead of answering the question the title asks. The phrase "a fifth separate case" refers to five directories the article never named. Two claims about how long a review takes are stated without a link to the source that says so. And the catalog facts, the ones worth citing most, were spread across six different sections with no single place to extract.
Our writing rules caught none of these, and the report that found them is honest about why. The rules measure shape: filler phrases, answer-first headings, the brief's promises against the finished text. Vagueness has no spelling, so a human or another model has to read for it. That split is the whole argument of lesson 9, and it shows up here as four sentences a machine could not use.
A table is the most citable asset a page can hold
The geo report named the missing comparison table as the single highest-value asset the page could have, and listed its columns: catalog, what it requires, verification, link type, review time and current status. The content report reached the same conclusion from its own angle, scoring the post 78 out of 100 for AI citation readiness and naming the absent table as its one format gap.
A table row is a self-contained unit of meaning, so a system that extracts one row keeps the context that makes the row true. The part worth noticing is that the data already existed, written out in prose across six sections of the same article. Collecting what you already wrote into a table is the cheapest high-value change on either report, and it takes about twenty minutes.
The direct answer belongs where a parser reads first
Change three of five was a forty-to-sixty word answer to the question the page is about, placed at the top. The report even wrote ours out as an example, built from facts already in the article: publish to the official registry with its two artefacts, list on the client catalogs, submit to the connectors directory, and set up a test account first because health checks fail without one.
seodraft has a field for this. The tldr is the answer a reader can lift whole, and the blog renders it as a block above the article body, which is why the post already carried one. The geo report measured the body's own first sixty words and found them narrating. Both readings are useful: put the answer in the stored field, and then make sure the first paragraph of the body answers too.
Server-rendered HTML is the floor, and we cleared it
The technical half of the score is a floor you either pass or sit under, and passing it is ordinary engineering. Our page returned its full 1,228-word body, its headings, its FAQ as a definition list, its byline and its time element in the initial HTML, with no JavaScript needed. The report checked this directly and recorded a cache hit on the edge.
Crawler access was measured the same way. The command fetched the post with the OAI-SearchBot and the Claude-SearchBot user agents and got 200 from both, and read the robots.txt to confirm no bot-specific rule exists. Nineteen of twenty points on that criterion came from facts you can reproduce with two requests, which is what makes that column worth trusting more than the ones above it.
Google says llms.txt is unnecessary for its own search
Google's AI optimization guide carries a myth-busting note saying llms.txt is unnecessary for Google Search and neither helps nor hurts there (developers.google.com). That settles the file's value for the largest answer surface anybody is writing for.
Our site publishes one anyway, and both reports describe it as present and well built: a product summary, pricing, a pointer to the dated citation page, and the free tools with their Markdown alternates. The geo report's only complaint is that it lists no blog posts, and it files that as optional and low priority. Keep the file if you have it, spend no afternoons on it, and put the time into passages.
The site around the page scored 82
A second geo reading ran over the whole site during the audit, and it scored 82 out of 100 on the same five dimensions: citability 20 of 25, structural readability 16 of 20, multi-modal 9 of 15, authority and brand 17 of 20, technical accessibility a full 20. The dated citation page at /ai-info carried most of the citability score, with a direct definition, a last-verified date and explicit guidance about what a model should avoid inventing about us.
Its top recommendations were question-phrased headings on the pages a model is most likely to read, more extractable text on a thin home page, and entity signals off the site. That last one is the honest weak spot: no Wikipedia, Reddit or YouTube presence was found for the brand, and the audit flags it as expected for something this new and worth building anyway.
What the report could not measure
Two sections of the geo report are a model's reading, and the report says so in both places. Platform readiness, the table that rates Google AI Overviews, AI Mode, ChatGPT Search and Perplexity as medium or medium-high, is marked qualitative, because measuring it needs a tool that watches answers over time and no such tool was connected. Brand mentions were skipped for the same reason, with only the on-page entity links observed.
Keep those sections out of any sentence you say to somebody else as a result. A readiness score grades how quotable your page is; measured visibility counts how often an answer engine used it, and the second one needs its own instrument. Lesson 13 is about the instruments, and about what a run honestly cannot see without them.
What to change on your own page this week
Three changes cover most of the gap, and our report ranked them in this order for a reason. Collect your repeated facts into one table with named columns. Write the forty-to-sixty word answer and put it where both a reader and a parser find it first. Link the source for every external claim, and name whatever your draft currently calls "a fifth case" or "several providers".
Then re-run the command on the same URL and compare the two scores. The number is a checklist, so a page that moves from 72 to 85 has removed real problems you can point at. For the evidence that makes those passages worth quoting in the first place, the evidence feature page shows how the bank behind them works.
On our post we made four of the five changes on 2026-10-03: the real publish date, a dated table with every catalog, links to the registry's own documentation with the actual server.json, and a paragraph that names only the reasons it can describe. The 40 to 60 word answer at the top is still pending, and the next run of /seo geo on the live page will show what the four changes moved.