Engineering Blog

Checking AI website suggestions before customers see them

Updated · Get1pct Engineering

Short answer

Before a customer sees an AI-generated suggestion, we check the target page, original wording, and available search data in code. Suggestions that pass appear with the before-and-after text and a reason for the change.

Separate generation from judgment

Our daily workflow generates suggestions from page content and search data, then checks the output format, target page, and original text in code. When target search data exists, it matches the page, query, impressions and clicks against that data and accepts average position within 0.05.

We check protected terms, duplicate suggestions for the same page and field, and title and description length. For example, a suggestion that cites a different impression count from the search data is withheld.

Look for additions and omissions

We check for numbers absent from the page, unfamiliar contact details or domains, and text that resembles an injected instruction. Advertising terms are checked against the page as well. A suggestion is dropped if screening changes its original text, replacement or reason.

If a rule detects a lost number, strength term or apparent company name, we try one repair and run the same checks again. A failed repair is withheld. These are pattern checks, not a complete understanding of every important fact.

Review the suggestion before applying it

The site owner reviews the wording in context and decides whether to apply it. A page can change between generation and approval, so the WordPress integration reads the current value before writing. See “Checking the change before writing back to WordPress” for that part of the workflow.

Trace a fictional suggestion

Suppose a fictional bicycle repair page says “We repair bicycles.” A proposal to change that to “We service bicycles.” can remain a candidate if its page, evidence and other checks also pass. Finding the original text alone does not approve it.

Suppose the matched search row has 100 impressions, 5 clicks and an average position of 6.0. A suggestion claiming 120 impressions is withheld. A replacement that adds an unknown number or a screened advertising term can also be withheld. These numbers are invented to explain the check; pattern matching cannot detect every misleading implication.

What to verify before deciding

Put the target URL, original wording, replacement and reason next to the current page. If the reason cites search data, check the same page, query and period in the report. A sentence may pass automated checks while omitting an important condition, price or service area.

If you cannot verify the claim, leave it unapplied and ask the person responsible for the service. Finish when you can identify the target and the evidence and have chosen to accept, revise or skip the suggestion.

Sources

View open roles