We Hill-Climbed This Blog's SEO With Our Own Agent. The Score Rose to 92. Search Hasn't Noticed Yet.

We ran our platform's own SEO agent on this blog, built a scorecard of the live site, and kept a change only if the score rose and nothing else got worse. Six rounds took the score from 77.3 to 92.4 out of 100, mostly on checks nobody has shown matter. A second test asked real agents 20 questions about the site: they answered 30 of 40 attempts correctly at the start and 32 after one change, which is close to the noise. AI search still found the site for 1 of 20 real questions, the same as before. Whether any of this brings readers won't be known until we measure again on Dec 14.

The site is small. Google sent it 23 visits in the 90 days before we started, 11 of them to one post. So this is a note on what you can measure about SEO, agent readability and LLM visibility on a small site, what our own agent got right and wrong, and what the experiment taught us about our platform.

How the climb works

A hill climb makes one change at a time and keeps it only if a measure improves. Ours had five steps:

1. Read the research and sort each claim by how strong the evidence is: documented by the search engine or crawler operator, measured in a study, or our guess.
2. Score the live site with a fixed list of pass-or-fail checks. Each check says which of those three tiers it rests on.
3. Pick changes, some from our own agent's audit of the site.
4. Ship them to the blog in a pull request, and score the live site again once the deploy is up.
5. Keep or revert. Keep only if the overall score rose and no check failed on more pages than before.

Prompt · Codex image generation, Nov 2026

Desktop version

A detailed production-quality diagram, landscape 3:2, in a risograph print style: two inks, deep navy and warm orange, on warm cream paper, light grain. Clean geometric shapes, straight orthogonal arrows, consistent line weight, generous spacing, one clean bold sans-serif typeface, large text that stays readable when shrunk. No title. Spell every label exactly as written, nothing added, no extra words. Layout, read left to right. Far left column, top to bottom, two navy-outlined boxes joined by a downward arrow: "Read the research" then "Score the live site". One arrow runs right from "Score the live site" into the top left box of a loop. The loop is a clockwise rectangle in the middle of the page. Top left box "Run our SEO agent", arrow right to top right box "Pick the best changes", arrow down to right middle box "Ship to the blog", arrow down to bottom right box "Score the live site again", arrow left to a large orange filled diamond at bottom left of the loop with the label "Better, and nothing worse?" fully inside it. Above the diamond, one short arrow labelled "Yes" goes up to a solid navy box "Keep it", and an arrow goes up from "Keep it" into "Run our SEO agent". Below the diamond, one short arrow labelled "No" goes down to a solid navy box "Revert it", and a thin line goes from "Revert it" left and up the outside of the loop, ending with an arrowhead in the left side of "Run our SEO agent". Far right column, separated from the loop by a tall dashed vertical line, a single box with a dashed navy outline: "Re-measure in weeks: rankings and citations". Orange fills only the diamond.

Phone version

A detailed production-quality diagram for a phone screen, portrait 9:16, in a risograph print style: two inks, deep navy and warm orange, on warm cream paper, light grain. Clean geometric shapes, straight orthogonal arrows, consistent line weight, generous spacing, one clean bold sans-serif typeface. Very large text that stays readable on a phone. No title. Spell every label exactly as written, nothing added, no extra words. Layout: one vertical column of navy-outlined boxes joined by short downward arrows, in this order: "Read the research", "Score the live site", "Run our SEO agent", "Pick the best changes", "Ship to the blog", "Score the live site again". Then a wide orange filled diamond with plenty of room around its label "Better, and nothing worse?", the label fully inside the diamond. Below the diamond two solid navy boxes side by side with cream text: left "Yes: keep it", right "No: revert it", each reached by a short arrow from the diamond. One line collects both boxes and runs up the right edge of the page, ending with an arrowhead pointing left into "Run our SEO agent". At the very bottom, below a dashed horizontal line, one box with a dashed navy outline: "Re-measure in weeks: rankings and citations". Orange fills only the diamond.

Round 3 shows every step, including the first reading after a deploy, which nearly fooled us.

What the research supports

We read the primary sources before changing anything. A lot of what's written about AEO and GEO doesn't hold up.

The studies mostly agree that authority counts for more than anything on the page, and authority builds slowly, off the site. We shipped llms.txt anyway, because it's cheap and helps when someone hands the site to an agent, but we didn't expect it to move citations.

The scorecard

The scorecard fetches every page in the sitemap and runs checks a script can decide on its own. Anything it can't measure stays unknown and out of the score, which includes field Core Web Vitals, because the site doesn't get enough Chrome visits to have them.

Crawler access went to 100 without any crawler gaining access it didn't already have. They could all read the site before. Naming each one turns an accident into a decision, but it isn't a gain in reach.

Each round

Three things nearly fooled us:

• Round 2's check that the markdown copy has every paragraph the page has. It guards against showing agents different text from readers, which comes close to what Google's spam policies call cloaking. Its first version compared word counts and failed 11 posts. Every missing paragraph turned out to be an image prompt, which the copy rightly leaves out, so we replaced it with a paragraph check.
• Round 3's first reading. It said 91.3, with the home page's Lighthouse score down from 0.94 to 0.89 and old descriptions on pages the round had just changed: the site was still refreshing cached pages. After that we measured twice after every deploy and judged speed on the median of five pages.
• Round 5's deploy, which failed three times on flaky builds while the site kept serving round 4. We caught it because the health page reports which commit is live, and discarded a reading taken too early.

Five things didn't move:

• Speed. After round 6, LCP was over 2.5 seconds on all five sample pages; the median went from 3,415 to 3,311 ms, within the noise. The slow element is the lede's text, and most of its time is render delay on a simulated slow phone. Our two CSS and font fixes didn't help.
• First paragraphs. Fourteen posts open on a scene rather than a number. Rewriting them to pass a crude check would make them worse to read.
• Sources. 25 of 40 posts still have fewer than three outside links. Most are about our own systems, and two passes found nothing honest to cite.
• www. It has no DNS record. The redirect waits for the owner to add one.
• AI search. The same 1 of 20 questions at the start, after round 4 and at the end.

A last reading with no site change came out at 93.3, entirely from Lighthouse. We report 92.4.

How it changed for a reader, Google, an AI search engine and an agent

The agent track

The scorecard lists what the site does. It can't tell you whether an agent sent to the site comes back with the right answer, so we tested that directly with 20 fixed questions: 13 about a number, 3 single facts, 2 that need two pages and 2 about the newest post. Each answer is checked against the post's own text. Real agents answered through our gateway, twice each:

• Navigator from llms.txt: Claude Haiku 5.5 with one tool, fetch, told to start at /llms.txt.
• Navigator from the home page: the same agent starting where a generic agent lands.
• Claude with web search: Anthropic's own search and fetch tools, limited to this site.
• GPT-5 search: searches the web but can't open a page it's given. Measured once.

Claude with web search moved by two between runs with no change it could see, and its two tries differed by two at the start. So a change of two is noise.

• Agent round 1, kept. A search page for agents: /search?q=words returns matching posts and paragraphs as markdown, and llms.txt links to it. The navigator from llms.txt went from 30 to 32, from 2.7 to 2.1 fetches a task, and from $0.096 to $0.068 for 40 attempts. The gain is at the edge of the noise, but it cost less and took fewer steps, so we kept it.
• Agent round 2, reverted. Round 1's search put the wrong post first for one question, so we ranked by the single paragraph matching most of the query. The right post came first, but the navigator went from 32 to 31, so by the rule it came out. The revert kept one sentence fix from the same pull request.

The replay shows the biggest problem the agent test found, and it wasn't on the blog. Our gateway's security scanner refused the navigator on 4 or 5 of the 20 tasks in each run. One post quotes real prompt-injection strings, so refusing it is the scanner working. But it also refused a post with the heading "The OpenRouter cost leak", and a table with the word "Leak" in it: the bare word was enough. Round 1 cut the refused tasks from 5 to 4. In round 2 the scanner refused the search page that held the right answer; the gateway doesn't record which rule fired.

Running our own agent on it

seo_research had one run on record, from May, for $11.46. It was built to outline a new page, and its reviewer rejected any report without keyword volumes from a named source. We have no Ahrefs or SEMrush, so in May both workers were sent back and no outline came out. This time:

The judge graded none of the three runs. The XP came only from the seat's own checks passing.

We also gave one Claude Sonnet 5.5 call the same brief plus nine pages from the live site, then had a fresh agent grade all three reports finding by finding, against the site as it was when each was written, without being told which was which.

None gave harmful advice. Run 2's top finding was the worst: it said our post titles were chart labels, having read a <title> inside an inline chart as the page's title. The real title comes first in the page for every crawler we tried. The single call found the most useful thing, the series sentence that read as "76 of 149 runs passed", and that sentence was ours. Three reports can't rank the approaches. They do show that every finding needs checking against the live page.

What we changed in the platform. #4043 has merged; the other two wait for review:

• #4044, the template. An audit mode with one worker, so a run fits $1 with room for review. The reviewer accepts "volumes unavailable" and rejects anything on Google's spam list. One required check fails a search volume with no named source, another requires the report to name the pages it checked, and the workers lose their shell, so text on a fetched page can't run a command. An independent review held the first version: an honest report could fail it, an invented number could pass, and a one-word report earned XP. Each is fixed with a test.
• #4043, the gateway. The blog org's key had no key_id, so every LLM call made with it failed with a 500.
• #4045, the scanner. Our first fix for the "leak" rule was too loose: review found about 45 attacks it would let through. The revised fix lets none of them through.
• The Office. The seat showed as "Seo Research", with a $6 bar where its level allows $1. The fix in #4044 uses the template's own name and shows the level's limit.

Two smaller ones: the gateway's cost header is an estimate (for GPT-5 search, about twice the recorded cost), and its PII scan refused our audit evidence because the About page shows a contact address.

In the Office

The seat sits on the floor of our Agent Office like the other agents. This is run 3 finishing, captured from the live Office for the blog's org:

The ask was to move the Office up a couple of levels through real work. That hasn't happened, and we didn't force it.

A run that meets its own objective earns 1 XP. The first run in a new seat earns nothing, because the seat's objective is adopted by an hourly sweep after it. A thumbs-up is worth 10, and that's the owner's to give. At one run a week the seat reaches level 2 in about 23 weeks. The blog org earned 34 XP in its last seven days, all from checks, which puts its level 3 around Nov 23. That date comes from the agents already running, not from this work. A weekly schedule for the seat was refused because our API key isn't an admin key, which is the right answer, so it's on the owner's list.

Picking this up

If you want to repeat this on your own site:

• The rule: one change at a time, measured on the live site, kept only if its measure rises and no check gets worse. Judge site changes on the scorecard and agent changes on an agent test, never the other way round.
• The noise: measure twice after every deploy, read Lighthouse as ±0.1 per page and judge speed on a median, and treat a change of 2 in 40 agent attempts as noise. Report each try, not only the total.
• The agent test: fixed questions with answers you can check against your own text, a mix of numbers, multi-step questions and "what's newest", run twice on at least two kinds of agent. Count blocks and errors separately from wrong answers.
• The scorecard: the seven categories in the scorecard table, each check tagged with its evidence, so a high score can't hide that most of it is guesswork.

Our next experiments, in order: rerun the same brief on the new template once it merges, and grade it blind against the old; retry the search ranking once the scanner fix lands; earn links by sharing the audit numbers where people write about evals; and measure Perplexity and Gemini, which we couldn't reach this time.

What we'll know on Dec 14

Almost everything that matters is still unproven. Rankings and citations take weeks to move, and so far we've only shown that the site does what our checks ask and that agents answer a little better. We wrote down in advance what would count as working: by Feb 1, 3 or more of the 20 questions on either model with web search, with the same question on two dates, and 50 or more Google visits per 90 days, across more than one post. If those numbers don't move, the likely reason is the one the research gave us at the start: other sites have to link here, and nothing on this one can do that.

We measure again on Dec 14 and Feb 1, with the same questions and the same scripts.