PageScan
A subscription product that crawls a website, checks every page with a language model, and tells the owner exactly what is wrong with their writing. Priced from what it actually costs to run.
The problem
Small organisations publish websites with typos, awkward grammar and inconsistent wording all over them, and nobody spots it, because the people closest to the copy have read it too many times. Professional proofreading is priced for publishers, and a twelve page business site can never justify it.
What we built
A product you can use without talking to anyone. Enter a domain and PageScan crawls it, pulls the readable content out of each page, analyses it in batches with a language model, and hands back a graded report with every error located and explained. Organisations sign up, subscribe, manage their sites, and scan again once they have made the fixes.
Behind that sits a crawler, a batched analysis pipeline that handles retries and back pressure, a results store, subscription billing, and an operator console for managing plans.
We measured what it cost before we priced it
The plans were not set by looking at what competitors charge. We instrumented the pipeline, ran the crawler over hundreds of real pages, and worked out the actual cost of analysing one. That turned up something we had not expected: pages get shorter the deeper you crawl, because the content-rich pages of a site are the ones you reach first.
That measurement turned the intuition upside down. The expensive act is starting a scan of a content-heavy site, and going deeper into one is cheap. So the plans ended up generous on page count and strict on scan count, which is the opposite of the obvious design, and worth roughly double the margin at the top plan.
A failure that looked exactly like success
A review of the analysis pipeline found that the model's answer could be cut short when a page produced more errors than the output budget allowed, and that the parser read a cut-off answer as a clean pass. The worst pages on a site were the ones most likely to come back graded as perfect.
Nobody had reported it, because a silent failure looks like success from the outside. We found it by working out the ratio between how many errors pages really contain and the configured output limit. The same review turned up an old input limit, set for a model that had since been retired, which meant roughly a third of the text on a site's most important pages was never checked at all.
Both were fixed in a deliberate order, because lifting the input cap first would have made the silent pass happen more often.
How we handled the data
- Scan limits are worked out and enforced on the server, from the signed-in session, because anything enforced in the browser can be edited by the person using it.
- Each organisation's data is separated at the database level, with the boundaries covered by an automated test suite that runs on every change.
- The crawler treats every page it fetches as hostile input, and the analysis path is capped so that one site cannot eat the budget for everybody else.
Outcome
PageScan is live, with self-service sign-up and paid plans. The more useful thing for anyone thinking about hiring us: this is a product where we did the unglamorous work of measuring what things really cost, and went looking for the failures that do not announce themselves. That habit comes with us into client work.