How it works
One layer measures. One layer acts.
Most SEO products stop at the report. The hard part was never knowing what is wrong — it is doing something about it, every week, without a person in the middle. This page shows exactly how that runs.
Why this is built the way it is
If publishing solved search, copywriters would own the internet.
There are millions of people who write well, and write constantly. Their own sites are not sitting at position one. That single observation kills the premise behind most of this category: that ranking is a volume problem, and enough words will eventually solve it.
It never was a volume problem. Search engines do not rank effort — they rank a set of specific, measurable things, and an answer engine rewards a different set again. A tool that only generates text is optimising the one variable that was never the constraint.
What search actually rewards
- A page that can be crawled, indexed and understood without friction
- One page per intent — not five competing for the same query
- Coverage of the entities the whole top ten treats as mandatory
- Internal links that tell the site which page is the answer
- Authority earned from domains that themselves have authority
What answer engines reward
- An entity a model can identify with confidence — who you are, stated plainly
- Claims with a source attached, because unsourced text is risky to quote
- Structure a machine can extract: definitions, tables, direct answers
- Corroboration elsewhere — a single site asserting something is not enough
- Consistency, so the same fact does not contradict itself across pages
Almost none of that is writing. It is diagnosis, structure, prioritisation and authority — which is why a general-purpose writing assistant cannot do this job, however good the prose is. Our models are never asked to “write an article”. They are handed a specific task on a specific URL, under constraints written by working SEOs, and everything they produce is scored against those two lists before it can go live.
| Technical + demand | Authority data | AI citations | Executes changes | Refuses bad output | |
|---|---|---|---|---|---|
| Classic SEO suites | yes | yes | no | no | n/a |
| AI writing / SEO autopilots | partly | no | no | yes | no |
| Answer-engine (GEO) tools | no | no | yes | no | n/a |
| RankMerit | yes | yes | yes | yes | yes |
We do not claim to be the only SEO product — there are dozens, and we measured 70 of them in our own index. The claim is narrower and checkable: no other product runs all five columns in a single loop. Diagnosis without execution is a report. Execution without diagnosis is a content firehose.
An X-ray of the site, not a checklist
Seoluma is the measurement half of the partnership. It crawls your pages, reads your Search Console, tracks your positions daily, and measures your authority against a web graph we build ourselves — 236.7 million domains and 7.12 billion links.
Every run produces five kinds of signal:
- Technical state — crawl errors, noindex accidents, broken canonicals, missing schema, thin and orphaned pages
- Real demand — search volumes straight from the Google Ads Keyword Planner API — not third-party estimates
- Actual performance — the queries you already get impressions for, and where you sit for them
- Authority gaps — which domains pass real authority in your niche, and which are recycled noise
- AI citation share — whether answer engines quote you or a competitor for your money questions
Findings become one thing: a ranked list of what to do next
A missing page, a broken canonical and a GEO gap look nothing alike, but they become the same kind of record — with a number attached. That number is not a guess:
priority = (expected gain × confidence) / effort
Confidence asks one question: is there a page in the top ten within our own authority band? If the first page belongs entirely to DR 80 giants, writing another article there is a waste — the system deprioritises it instead of burning your month on it.
There is also a hard lock: if a page already targets that intent, a new page cannot be created. The item is converted into a revision of the existing one. That single rule is what stops a site from ending up with 300 posts and no traffic.
A finding is not an instruction. We turn one into the other.
You can export any audit and paste it into a chatbot. What comes back is generic advice about a site the model has never measured, judged against an algorithm it last saw during training.
RankMerit does something different with the same raw finding. It writes a task: which URL, what to change, under which constraints, and against which measurement — versioned to the current state of the algorithm, not to what was true last year.
The models are driven by rules written by working SEOs: what to touch, in what order, and — just as important — what to leave alone.
Connect up to three AI providers. We drive them.
Already paying for OpenAI, Anthropic or Gemini? Connect your keys and the platform costs $9.90 a month. Generation runs on your account, with your models, at your cost — and every instruction we send is logged, so you can see exactly what was asked and what came back.
This is also a cost lever: heavier models for the hard structural work, faster and cheaper ones for routine fixes.
Change classes you approve are applied automatically. Everything else waits in a queue for one click. Every change is reversible from the log — the previous value is stored before anything is written.
Two layers whose only job is to stop work from shipping
Layer one is mechanical — no AI involved. An unsourced statistic, an intent that collides with an existing page, missing entities that the whole top ten covers, no internal link in or out, template sentences repeating across the site. Any one of these fails and nothing ships.
Layer two is a separate judge model with an adversarial brief: find a reason not to publish. The model that wrote the draft never scores it. 80 or above ships. Below that it is rewritten, and the exact rule it failed is fed back into the rewrite.
If it fails twice, it goes to a human queue — not to your live site. This is the deliberate opposite of the tools that publish something every single day regardless of whether it deserves to exist.
Every change is checked against what actually happened
Each action is stamped and re-measured at 14 and 45 days against impressions, clicks, position and AI citations.
- It moved — the item closes and that action type gains weight for your sector
- It did nothing — the same approach is not retried — the item escalates to a heavier action
- It got worse — automatic rollback
Those weights are kept per sector, so the system learns that schema work moves the needle in one vertical while meta rewrites do not. That feedback loop is the part a report cannot have.
Boundaries
What this system will never do
A method is defined as much by its refusals as by its features. These are permanent, not temporary limitations:
✕Buy or place links
Paid schemes violate Google's link spam policy and are cheap to detect. We build assets worth citing instead.
✕Blast directories or run a PBN
Volume without editorial intent is ignored at best, penalised at worst.
✕Publish on a schedule
Content is produced against demand in your Search Console, never to fill a calendar.
✕Fabricate proof
No invented review markup, no unsourced statistics — the mechanical gate rejects both.
✕Manage your social feeds
Video and social are measured as search surfaces, not run as channels. That is a different product.
✕Hide what it did
Every directive, change and rollback is logged and visible to you.
See the first step run on your own site.
The diagnosis layer is live and free to try. Enter a domain and it crawls your pages, pulls real search volumes and shows you what it found — in under a minute.
Scan my site — free