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. It covers the audit, the plan and the homepage mockup, because those are what a report shows. What a model writes for businesses that pay us is further down, and unlike either of those, it is published.

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 of the audit 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.

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

A 90-day plan and a homepage mockup some people are shown, and — for businesses that pay us — replies to Google reviews, posts on their Google profile, the words in their monthly report and the pages of a website they buy. They differ in the way that matters most: the plan and the mockup are on a page, and a reply, a post or a website is published where anybody can read it.

The plan is a rewritten Google Business Profile description, the categories to set, four posts and a review-request message. A model writes that. The wording is its own, which is the point: it is a draft of the work, written for one business rather than picked off a shelf.

• Every fact in it has to come from somewhere — The services, the places and the trade it names all come from the same scan the report came from. Before a plan is shown, it is checked against those facts one claim at a time — a service we never found, a town we never read, a number nobody measured. A plan that fails is thrown away rather than patched, and we would rather show nothing than something we cannot point at.
• It never carries a price, an offer or a guarantee — Those would commit a business to something in front of its own market, and we are not in a position to do that on anybody’s behalf. The review message offers nothing in return for a review either — that breaches Google’s policy and the profile at risk would be theirs.
• The categories are chosen, not invented — A model asked to name a Google category returns something that sounds right and cannot be set. So it picks from Google’s own list, and an answer outside that list is rejected.
• Its posts are published only if the business buys — To somebody who has not bought anything it is text on a page, and it touches no Google profile — we hold access to none until a business pays us to look after theirs. If they do buy, its four posts become their first month, published on their profile the way every post below is.

Some reports come with a homepage made again from the business’s own site, shown beside what that site looks like on a phone today. A model writes its words — a headline when the site has none, a line about each service, a paragraph about the business, and the words on two buttons. Everything else on it is ours or theirs, and none of it is the model’s:

• The facts on it are Google’s, and the model never sees them — The phone number, the rating and review count, the opening hours and the town are drawn from the business’s Google listing by our own page. The model is not given them, so it cannot repeat one as its own.
• Every claim in its words has to come from somewhere — The services and places it names come from the same scan the report came from, and it may not quote a customer, say how long the business has traded, claim a licence, insurance or an award, make a guarantee, name a price or a person, or promise emergency or round-the-clock service the site does not list. A page that breaks one of those is thrown away, and the model gets one more try before there is no mockup.
• The logo, colours and typefaces are the business’s own — Read off their site by code, never by a model, fetched once and shown from our site rather than loaded from theirs. A colour too pale or too dark to read on is made readable; a business whose site gave none gets its trade’s. A line under the mockup says which parts came from the site.
• The picture of the site today is Google’s — It is the screenshot Google’s own speed test took while loading the site on a phone, not a picture we made.
• Nothing the model writes is put on the page as code — It answers in a fixed shape, and our page draws it. That is why a mockup looks like the same few sections in a different business’s colours, and it is the point: a page a model wrote the code for is a page we would be hosting on its word.
• It is illustrative, and nothing about it is live — The page says so — “Illustrative — not your site, and not a guarantee” — and the words are ours, not the model’s. We do not touch the business’s site, and nothing on the mockup is published anywhere.

A business that buys a website from us gets its pages’ words written by a model — every page, from the details and services the business confirmed, the headline on its current site, and anything it asked us to avoid or change. The business approves the website before anything of it is live, and it does not say a model wrote it.

• The facts on it are the business’s own, and the model never sees them — The phone number, the address, the hours, the rating and the number of reviews are drawn by our pages from what the business confirmed and from its Google listing. The model is not given them, so it cannot get them wrong.
• Every page is checked before anybody sees it — The mockup’s checks, on every page: a service or a place it was not given, a year, a licence, a price, an offer, a guarantee, a customer’s words or a person’s name, and the words are thrown away. The model gets one more try, and then there is no website yet rather than one we cannot point at.
• Nothing the model writes is put on the page as code — It answers in a fixed shape, and our templates draw it — one of six designs the business chooses between. That is also why switching the design changes no words.
• A person reads it before the business is asked to approve it — We read every page before we send the preview, and a change the business asks for is made by a person or written again with their request in front of the model — and checked again.

A business that pays us to look after its Google profile gets one post a week. The first month is the four posts from its plan, the ones it read before it bought. After that, a model writes each month’s posts, from the same facts a plan is written from — the trade, the services and the places we found, the month the posts are for — and the titles of the last few posts, so it does not repeat itself.

A person publishes each post by hand, on the business’s own profile, under the business’s name. The model never touches the profile, and a post does not say a model wrote it.

• Every fact in a post has to come from somewhere — The same checks as the plan: a service we never found, a place we never read, or a number nobody measured, and the month is thrown away. The model gets one more try, and then there are no posts that month rather than posts we cannot point at.
• A post never carries a price, an offer or a guarantee — and a second model checks — A fixed check refuses prices, discounts, guarantees and deadlines outright. What survives is then read by a second, smaller model that asks one question of each post: does this commit the business to something a customer could hold it to? A post it says yes to — or that it cannot give an answer about — waits for the business to approve it, however the business has chosen to handle posts. That model can only make a post wait; it can never send one out.
• The business decides how posts go out — It can approve every post, or let a post go ahead three business days after we tell it one is ready if it has said nothing. The second only starts once the business has approved a first batch by hand, and every email about a batch says the date it goes up. A business can decline any post, and a declined post is never published.

A business that pays us gets a report each month: where it shows in Google’s top three across its area, the reviews it got, the replies and posts that went up, and how often its team asked customers for a review. We measure every number; a model writes a few sentences around them — what moved, and what we are doing next.

• The model never works out a number — It is handed every figure it may use, including the changes from last month, which we subtract ourselves. If its words carry any other number, or mention calls, directions or website visits — which we do not measure — they are thrown away.
• A person reads every report before it is sent — We approve the words, have them written again, or send the report without them. A report nobody has approved by the fifth of the month goes out with its numbers and without the model’s words.

A business that pays us to look after its Google profile can have us answer its reviews. A model drafts each reply, from that one review — the reviewer’s public name and what they wrote — the tone the business picked, a few replies the owner wrote themselves if they gave us some, and the last few replies written for the same business, so it does not say the same thing twice. It is told nothing else about the business, not even its name.

We then publish the reply on the business’s own Google profile, under the business’s name. A person does that, by hand. The model never touches the profile, and the published reply does not say a model drafted it — it is the business’s reply, and the business decides what may go out under its name:

• A reply to a complaint is never published without the business approving it — Three stars or fewer. Until the business says yes it stays a draft, and a draft nobody approves within three business days is never published. A business can also choose to take our draft and post their own words, or to have us stay out of complaints entirely.
• A reply to praise can go out without the business reading it first — Only if they chose that. They can also ask to approve those too, or to have us leave praise alone.
• Some reviews never get a reply from us — Anything alleging damage or injury, fraud or overcharging, a named employee’s conduct, legal action, or a safety problem is held for the business owner, and no reply is drafted at all — and when the step that looks for those cannot give an answer, the review is held the same way. Nor is a reply drafted to a rating with no words, a review the business has already answered, or one that was posted more than a week before we first saw it.
• Every draft is checked before anybody sees it — A reply may not admit fault, retell what a complaint says happened, promise a refund or a return visit, offer a price or a discount, name anybody on a complaint but the reviewer, mention a name, a number or a way to get in touch that is not in the review, or ask for a review to be changed. A draft that breaks one of those is thrown away, and the model gets one more try before the review is left for a person.
• The person publishing it checks it is the right review — They confirm the name on the review they replied to, and a name that does not match is refused before anything is recorded.

• 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. That is true of the report whether or not a plan was written for you — they are separate steps and the plan is never shown a finding.
• 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.

It has already happened five times. Until 14 September 2026 this page said one step of the product was a model and everything else was computed. The 90-day plan made that false on the day it shipped, so the sentence changed in the same release and the date at the top of this page moved — which is the version your privacy choices are recorded against, so you were asked to look again.

Until 17 September 2026 it said that only one thing outside the audit was written by a model, because we published nothing a model wrote. The review replies above changed that on the day we could first publish one, and the page and its date changed with them.

Until 24 September 2026 it said that nothing in a plan was published by us. The posts above changed that on the day a paying business’s first month could be written, and the page and its date changed with them.

Until 25 September 2026 it said nothing of the monthly report. A model writes the few sentences around its numbers, and a person says yes before any are sent; the page and its date changed on the day the first one could be drafted.

Until 28 September 2026 it said nothing of the homepage mockup, and the sentence under every report counted two parts that are AI rather than three. A model writes the mockup’s words, and the page, the sentence and the date changed on the day the first one could be made.

We still 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. Neither ships yet. When either does, it happens the same way.

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