Skip to content

Three numbers you can watch move

Canopus is an autonomous growth manager that audits a SaaS site on three axes, search, citability by AI answer engines and conversion, and scores each one out of 100 from deterministic rules.

The audit runs on its own and the scores come from weighted checks over what was actually on the page. A model reads the same pages and adds what a checklist cannot see, and it never touches the number. That is what makes "41 on the first run, 68 on the third" a fact rather than a reworded prompt. On Growth and Scale, the findings are what the SEO and GEO content agent writes from.

What happens when you run an audit?

  1. Your pages are read once

    The landing page and the pages it links to are crawled behind an anti-SSRF guard, up to twelve pages. Titles, headings, buttons, links, images, JSON-LD and the visible text are stored. All three audits read that same crawl, so the three numbers describe the site on one day rather than on three.

  2. The rules score the site

    Fourteen checks for search, eight for AI citability, ten for conversion. Each one is a yes or a no with a weight attached, so a missing title tag costs more than one image without alt text. The result is rounded and clamped to 0 to 100.

  3. A model adds judgement

    It is given the same pages, your product brief, your validated customer profile and the rule findings as observations, so it does not repeat them. It adds at most eight findings and writes a two or three sentence summary. When it can, it writes a rewrite: a headline you can paste, not a description of a better headline.

  4. The run is kept, never overwritten

    Each run writes one row per audit type, grouped under a run id so the three belong together. Rows are never updated. A run that happened without a model is marked as such, so a rules-only score is never silently compared to a full one.

What does this replace?

  • Running a site through three different scanners and reconciling three reports

    One crawl, three scores, one ordered list of findings on one screen.

    Replaces: A desktop crawler
  • Pasting your landing page into a chat window and asking what is wrong with it

    A model that has your product brief and your validated customer, and that is told not to repeat what the rules already measured.

    Replaces: A chat window
  • Guessing whether AI assistants can describe your product at all

    Eight checks for citability: the definition sentence, question headings, FAQPage markup, entity markup, comparison pages, pages long enough to quote.

  • Keeping a spreadsheet of what you fixed and when

    Every run is kept with its score, its findings and the number of pages it considered.

    Replaces: A spreadsheet
  • A one-off audit from a freelance consultant

    The same checks every time, so the second run is comparable to the first.

How is a score calculated, and what happens to the findings?

The search audit asks the questions a crawler asks. Is there a title, and is it between 15 and 60 characters, which is roughly what a result shows. Is there a meta description between 70 and 160 characters. Is there exactly one H1, and at least two H2s. Are there at least 300 words in the HTML, because below that there is nothing to rank, and below 80 words the page is almost certainly drawn in the browser and invisible to anything that cannot run scripts. Is there a canonical URL, an og:title and an og:image, any JSON-LD at all, at least five internal links. Does any page take more than 1500 ms to return its HTML. Do several pages share a title. Is there a pricing page, and does any linked page answer with an error. Fourteen checks, each weighted, each deterministic: it costs nothing to run, so it runs on every page rather than on a sample, and a founder who fixes a title sees the score move by a predictable amount.

The AI answers audit, what the industry calls GEO, asks a different question, and it is the one worth reading. Search ranking is about being found by someone who is already searching. Citability is about being quoted when someone asks an assistant what they should use to do a job. The checks are for the properties that make a page liftable. Does the opening copy contain a sentence of the form "X is a ...", computed over the whole page rather than the truncated excerpt. Are at least three headings across the site phrased as questions, where the question word has to start the heading or the heading has to end in a question mark. Is there FAQPage JSON-LD. Are there at least two of Organization, SoftwareApplication, Product, FAQPage, HowTo, Article or WebSite, so an assistant can tell what kind of thing this is and who makes it. Is the product name in the title or the H1. Is there a comparison or alternatives page, and does any page name the tools a buyer is also considering. Are there at least three pages over 400 words, because assistants cite pages that explain rather than pages that pitch. Sites that rank well in search fail this badly, and their owners never see a search result to be missing from.

The conversion audit measures the part of conversion that is not a matter of taste. Is there a call to action at all. Does one of them go straight to a sign-up, a trial or a get-started page. Does the first one appear within the opening 700 characters of copy. Do the buttons say what happens next, or do they all read "Learn more". Are there more than twelve distinct actions, which makes the page a menu. Is the headline at least 20 characters, and is it something other than the product name. Is there any proof on the page: a customer count, a named customer, a testimonial. Is pricing linked, and does the pricing page have at least 120 words of readable text. A check that cannot be asked is removed from the denominator rather than counted as passed, which is why a site with no pricing page earns no points for not having a thin one. The audit reads structure and copy rather than a rendered page, so it takes no screenshots and has no opinion on your layout or your contrast; adding visual criteria would change the rule weights and break the comparability of every score already recorded, which is why it is not bolted on quietly.

Every finding carries a severity, a detail and a recommendation, and often a rewrite and the URL it is about. Severity measures lost revenue, not untidiness: high means it is plausibly costing customers today. The list is sorted high first, because you are meant to work down it. The score itself is a weighted result of the deterministic rules only, and the model never changes it. This is a deliberate constraint, and it has a cost: changing a rule weight moves the history, so a weight change has to be treated the way a schema migration is, with a new series rather than a continuous curve. The crawl is capped at twelve pages, which is the right size for a marketing site and the wrong size for a large documentation site, and every run records how many pages it considered so a score is read in context.

The findings are not the end of the loop. On Growth and Scale, they are the input of the SEO and GEO content agent: from the latest audit, it proposes an editorial plan where each piece answers one finding, as a comparison page, an alternatives page, a how-to, a definition with FAQ or a teardown series, and it drafts the pieces you keep. Beside the plan sits a kit to paste, built from your validated brief and the scanned pages: an llms.txt, the structured data and the title and description of your home page. You edit, publish on your own site, then ask for a rescan: the site is read and audited again, and the scores are compared with the audit the plan was drawn from, with the findings resolved since. The history on this page becomes a before and after of work that was actually done rather than a number that drifted. Either way, the findings are yours to work down, and the rewrites are already written to be pasted.

Questions founders ask

See your own numbers first

Connect Stripe and your real revenue is on screen in five minutes, before an agent has said a word. Create your account, answer the questions, and your strategy gets written.