AI disclosure

That sentence appears under the findings on every report, because a disclosure that only exists on a page nobody clicked has not disclosed anything.

The audit runs a sequence of steps. All but one of them are ordinary code: fetching a Google profile, reading a page’s source, checking a performance score, looking up a map position, counting reviews.

One step is a model. It reads the text of the website and answers a single question: what trade is this business in? Plumbing, HVAC, roofing, landscaping, and so on. That answer does two things you can see:

• It picks the search terms — The map positions we report are measured against terms built from that trade. Classify a roofer as a landscaper and the whole ranking section is measuring the wrong thing.
• It decides whether one finding applies — We compare the trade the site reads as against the category Google has the profile filed under. When Google’s category is one of its catch-all ones and the site is clearly a trade, that is a finding. Without the classification there is nothing to compare and the check does not run at all.

That is the whole of it. If the trade at the top of a report is wrong, assume the ranking section is measuring the wrong terms, and tell us.

• It does not write the findings — Every finding’s title and explanation is written by a person and kept in a fixed list. The model has never seen that list and cannot add to it, reword it, or choose which entry appears.
• It does not set severity — How bad a problem is comes from a rule applied to a measured fact. A model that could promote a finding it found rhetorically useful is exactly the failure this is built to prevent.
• It does not decide what is shown or withheld — What appears before the email lock is the worst findings by severity, computed. The locked ones are real findings with their words removed — never invented rows.
• It does not touch pricing or qualification — Prices come from a rate card. Our internal view of whether a business is a fit for us is computed by code reading versioned rules. The model sees neither.
• It does not read anything private — It is given the public text of a public web page, stripped of markup and scripts before it ever gets there. It is never given your email address.

Classification fails on businesses that describe themselves oddly, on sites with almost no text, and on companies that genuinely do several trades. When it cannot place a business it says so rather than guessing, and the findings that depend on it are marked as not measured instead of being filled in.

• The page is sanitised before the model sees it — A stranger’s web page can contain text written to manipulate a model reading it. Ours is stripped and labelled as data before the call, and the answer is checked against a fixed schema rather than trusted.
• A bad answer is rejected, not patched — An answer that does not fit the schema is thrown away and the step records that it failed. It is never coerced into something that looks valid.
• Every call is logged and tested — Model, prompt version, tokens and cost are recorded on every call, and a suite of fixed examples runs on every change so the classifier cannot quietly get worse.

We intend to have a model write more of a report than it does today — a narrative summary, and a short list of what to fix first. None of that ships yet. When it does, this page changes in the same release, and because the date at the top of it is the version your privacy choices are recorded against, you will be asked to look again.

The privacy policy covers what is sent to the company that runs the model. Anything else: privacy@alltradesdigital.com.