The SEO and AI visibility agent finds gaps in Google rankings and AI answers, researches primary sources, and drafts a small fix with a staging preview. A person approves the exact change before it ships. The agent measures the next 28 days against similar unchanged pages.
Inside the SEO & AI Visibility Agent
Evidence enters the loop. A reviewed change leaves it. Measurement decides what to propose next.
A search or citation gap becomes a sourced proposal and a draft on an isolated runner.
The reviewer checks sources, changes and screenshots. Only the exact version they approved can merge.
Compare 28 days before and after against similar unchanged pages. A verdict may be better, worse or inconclusive.
The diagram shows the measurement method. It does not show measured ranking or citation gains.
A buyer looking for a supplier today asks two places. One is still Google’s list of ten links. The other is an answer — from ChatGPT search, Claude, Perplexity or the AI Overview at the top of Google — assembled from a handful of pages the engine decided to trust and, sometimes, cite. Being on page two of the first is invisible. Being left out of the second is worse, because nobody is shown a list to scroll.
The work that improves both is not mysterious, it is just relentless. Find the query a page nearly ranks for. Find the question an engine answers from someone else’s page. Read why: the answer is buried under preamble, the fact has no source, the definition is vague, the page has never been fetched by an AI crawler at all. Write the smallest change that fixes it. Then wait a month and check, honestly, whether it worked — which is the step almost nobody does.
That loop is a week of an expert’s time per handful of pages, forever. So most sites do it once, in a burst, from an audit tool’s list of warnings, and never measure a single change. This agent runs the loop every week instead, and it is strict about the two parts that make it trustworthy: every change is sourced, and every change is approved by a person before anyone outside sees it.
search · AI answers · crawlers
a gap, with evidence
primary sources only
isolated runner · allowlist
a person merges it
28 days vs control pages
Search Console says which queries a page appears for and where it ranks. Weekly citation probes ask the connected AI search engines the questions a buyer actually types and record whether the answer cites the page that should be cited. Server-side counts show which AI crawlers fetched which pages, and a radar of releases, papers and repositories shows what the field is talking about this week.
A query on page two, a question answered from a competitor’s page, a section that opens with three sentences of preamble before the answer, a page an AI crawler has never fetched. Each becomes a proposal that carries its evidence and a plan, so the reason for a change is on the record before anyone writes a word of it.
Research runs with web tools and nothing else. It can cite only pages it actually fetched or a search actually returned, and a new number or year in a draft must already appear in the page or in one of those sources. Thin or failed research stops the job there rather than letting the model fill the gap.
The draft is written on a GitHub Actions runner that may edit only the files the job names. A separate verify step with no secrets and no open internet checks the change, renders the page, runs its tests and takes before-and-after screenshots on desktop and phone. The result is a pull request and a private staging preview — never a change to the live site.
The reviewer sees the evidence, the sources, the diff and the screenshots, ticks that the facts were checked, and approves, sends it back with notes, or rejects it with a reason. The merge is of the exact version reviewed: if a single byte changed after review, it does not merge. Search engines are then told the page changed.
Each change is judged on outcomes — clicks, impressions, position, AI crawler fetches and citations, enquiries — over the 28 days after it shipped against the 28 before, compared with three to fifteen similar pages that did not change. Most verdicts at modest traffic come back inconclusive, and the agent says so. A change judged worse is offered back as a restore, and a person decides.
An SEO tool tells you what is wrong with a page by its own rules. This agent starts from what the outside world did with the page — the queries it showed for, the answers that left it out, the crawlers that skipped it — and works backwards to the change. Several of these inputs come from the search and AI analytics agent, which counts and audits; this one acts on what it finds.
Generative engine optimisation sounds like a new discipline. Most of it is the old one done with more discipline. An AI engine assembling an answer pulls passages, not pages, so the changes that help are the ones that make a passage easy to lift and safe to attribute.
The answer goes first. A section that opens with its answer in one or two sentences, and explains afterwards, is quotable. One that opens with context is not.
Facts carry their source. A number with a date and a link is something an engine can repeat with confidence. The agent will not add a figure it cannot source, which is also the rule that keeps it honest.
Questions are asked in the reader’s words. A question-and-answer block written the way a buyer types the question is the passage most often lifted whole.
The page is reachable. A clear canonical, indexable text, structured data that matches the page, and no firewall rule quietly turning search crawlers away. Our llms.txt summary is an additional navigation aid, not a requirement for Google’s AI features or a guarantee of citation. See Google’s published AI search requirements.
Mechanism, not a promise. These are the changes most likely to make a page the better source; whether an engine cites it is still the engine’s decision, which is why every change is measured.
An agent that can edit a public website is only as trustworthy as the limits it runs inside, so the limits are enforced in code rather than written into a prompt and hoped for. The model only ever edits content. Redirects, links, registry entries and restores are mechanical operations, and the live site re-checks that a draft changed exactly what it declared.
The drafting runner holds none of the keys it does not need and cannot reach the production database. The step that checks a draft has no secrets and no open internet. Only the step that opens the pull request can push, and it never runs model output. Approving a draft merges the exact version reviewed: if it changed after review by so much as a byte, the merge is refused.
And it is slow on purpose. A few changes a week, one change per page, then a month of measurement before that page is touched again. That pace is what makes a verdict mean something.
The agent runs on QQuantum.ai, the site you are reading. It researches and drafts on GitHub Actions against this repository, previews on a private staging copy, and waits for an editor here to approve. The first change it drafted was approved and merged on 21 October 2026: an update to our Mistral model page after Mistral released its time-series models.
We have not published any figure for what it has done to rankings or citations, because the honest answer at this stage is that the measurement windows are still open. When a verdict exists it will appear here with its date and method, as every figure on this site does.
The agent cannot guarantee a ranking or a citation, and no honest system can: Google and the AI engines decide what they show. What it changes is how quickly a weak page is noticed, how carefully the fix is sourced, and whether anyone ever finds out if the fix worked. It publishes nothing a person has not approved, it does not buy links or spend on ads, and it never touches legal, pricing or contact pages.
Every page in this portfolio carries one of these. The claim a page refuses to make is the reason to believe the ones it does — see our work.
Chosen for this agent. Model selection is per task and reviewed as the families move — see /technology/models.
It is an AI agent that does the ongoing work of search optimisation and generative engine optimisation together: it finds the pages that rank just short of page one or are left out of AI answers, works out why from the data, drafts a small, sourced change, and measures whether the change helped. Ours runs that loop every week on our own site, and a person approves every change before it is published.
By asking them. Each week it puts the questions a buyer would ask to the connected AI search engines and records which pages they cite. Where the answer cites someone else, it proposes the change most likely to make the right page the better source: an answer in the first sentence of the section, a clearer definition, a fact with its source, a question and answer a model can lift. Whether an engine then cites the page is the engine’s decision, so the agent measures it rather than promising it.
No. Every content change is drafted, checked and previewed on a staging copy, then waits for a person. The reviewer sees the evidence, the sources, the diff and before-and-after screenshots, and approves, sends it back or rejects it. Only the exact version that was reviewed can be merged, and restoring an older version is always a person’s decision too.
Each change is compared over the 28 days after it shipped against the 28 days before, alongside three to fifteen similar pages that did not change in either window. It looks at outcomes — clicks, impressions, position, AI crawler fetches, citations and enquiries — not audit scores. At modest traffic most verdicts are inconclusive, and it says so and estimates how many more weeks it would need. A change judged worse becomes a restore proposal.
It is built to do the opposite. It is limited to three page changes and two articles a week, no change may rewrite more than a third of a section, every new number must come from a fetched source, and it writes in the site’s own voice rather than trying to look human. Articles carry a named editor who approved them. A small, sourced change you can measure is worth more than a hundred pages you cannot.
On our own site the paid APIs it uses — AI engines for the citation checks, keyword research and image rendering — run under a hard monthly cap of €75 with a separate limit for each provider, and a provider that reaches its limit pauses until the next month. Drafting runs on a coding-agent subscription and GitHub Actions minutes. A client deployment is scoped and priced per site.
Yes, as a managed engagement. It was built for our own site first, and the parts that matter — the evidence, the source rules, the review gate, the measurement against control pages — carry across. The part that changes per site is how a draft is written into your stack, so we scope that first. You keep approval of every change.
What we built, what stage it is at, and a link to the live thing wherever there is one.
yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.
Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.
Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.
Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.
Tell us which pages matter and which questions your buyers ask. We will show you where you are losing in search and in AI answers, and what the agent would propose first.
30-minute call · no deck · we will say if the answer is no