Why the generator is not allowed to invent facts

Language models are good at sounding right. For advertising a used car, that is the problem: “sounding right” produces heated seats that were never installed, a warranty nobody offered, and “award-winning” from no award. A dealer who publishes that ad owns the consequence.

So the first rule in Automotive.ad is not about tone or length. It is: use only the facts provided, and leave out anything that was not.

Structure before prompting

The generator does not receive free text. It receives a validated request — a vehicle block, an offer block, a dealership block — where every value came from a labelled form field. The instructions tell the model, in plain words, that these blocks are the only source of facts, that a missing fact is left out rather than guessed, and that the content of the blocks is data, not instructions to follow.

This matters more than any clever phrasing. When the input is structured, “don’t invent” becomes checkable: if the output mentions a feature, we can look for it in the highlights.

Guidance per part, not per prompt

Each part of each format carries its own guidance: a Facebook primary text leads with the price or the offer and holds no call to action; Google’s second headline is the price when there is one; a listing description opens with the vehicle and its strongest fact, never a question. The model gets these rules alongside the character limits, so it knows what each field is for instead of filling all of them with the same sentence.

Disclaimers are copied, not paraphrased

Financing language and dealer disclaimers exist because someone, somewhere, required exact wording. The generator is told to reproduce disclaimers verbatim in every format long enough to hold them, and to keep conditions such as “on approved credit” attached to the terms they qualify. It is not allowed to make them friendlier.

Length is verified, not trusted

Every part comes back through a validator that measures it against the platform’s limit. If something runs long, the generator gets one chance to shorten exactly those parts; anything still over is flagged in the result rather than silently accepted. Structured output from the model is treated the same way any external input is: checked before it reaches you.

What this costs

Copy that only uses the facts you entered is sometimes shorter than you expected. That is the trade we chose. An ad with three true sentences publishes; an ad with five sentences and one invented feature does not.

You can see the result of these rules on the examples page: one fictional dealership, one vehicle, eight formats, no edits.