The short answer: a marketing tool that shows you “£3,520 of opportunity” has usually added together things that are not the same kind of thing. Money you could stop wasting, conversions you might gain, organic clicks a page could earn, and a warning that a signal has stopped are four different objects. Two of them can be turned into pounds if you state an assumption; two of them cannot. We tried the single number, saw what it did to the reading, and removed it. Here is the reasoning, because it applies to any report you are shown.
What went wrong with the total
The first version of the HighRegard opportunity feed carried a headline: the sum of every estimate on the list. It looked decisive. It was also false precision. Underneath it sat a saving from negative keywords (a real, bounded amount), a projected number of extra conversions (dependent on an assumed value per lead), an estimate of organic clicks (dependent on a click-through rate the page does not have yet), and several items whose benefit was described in words: ranking stability, bidding quality, reach without extra budget. Those last ones had been given a nominal figure so that the ranking formula would work. The total added the nominal figures too.
The number was not a lie about any single item. It was a lie about the list, because it implied the items were commensurable and independent, and they were neither. Two recommendations about the same query overlap; fixing the tracking changes the estimate on every paid item; a click is not a customer.
What we show now
Every recommendation carries a kind: a saving, conversions, organic clicks, a risk, or a quality improvement. The feed totals each kind separately and never across kinds. So the top of the page reads something like: recoverable paid spend £525, potential conversions 62, potential organic clicks 259, two data or tracking risks, six quality improvements, four items blocked until tracking is fixed. You can disagree with any of those on its own terms.
Where a conversion estimate is expressed in money, the assumption is printed on the item: the account’s actual cost per conversion where one exists, or £45 per lead event where it does not, labelled as an assumption until the booking system is connected and a real value replaces it. Estimates that fall below a materiality threshold (fewer than five organic clicks, less than £20 in a window) are dropped before they reach the list, because a line that says “+0 clicks” is noise dressed as advice.
Observed, estimated, judgement
The same discipline runs inside each recommendation. What was measured is labelled observed: “£374.54 spent on 13 search terms with no conversions”. What was projected from it is labelled estimated, with its assumption: “≈ £300 per 28 days redirected, assuming 80% of the flagged spend is avoidable”. What we advise is labelled judgement: “add the confirmed candidates as phrase-match negatives”. If the judgement is wrong for your business, you can say so without arguing about the observation.
Why the AI never types a number
The overview includes a short written assessment. An earlier version let the language model format the figures itself, and it did what language models do with numbers when they are handed raw values: it wrote £1,741 of spend as “£1.74m”. The fix was structural rather than a better prompt. The model now receives a table of named values already formatted by the application, and may only write them as tokens. The application inserts the figures after generation, moves the citations into an evidence line under each paragraph, and validates the result: no unresolved token, no figure that is not in the table, no scaled unit, no paragraph without evidence. If any test fails, a rules-based summary is shown instead. The customer never sees the failure; the operator sees why it happened.
What to ask of any report
When a proposal or a dashboard gives you one number for “opportunity”, ask three questions. What kinds of thing were added together? Which assumption turned each of them into money, and is it written down? And which items depend on another item being true first? If the answers are not on the page, the number is a feeling, not a measurement.
The opportunity feed described here runs on synthetic data in the live demo; the rules the engine keeps are set out on the platform page.
Next step: if this raised questions about your own visibility, there are three honest ways to answer them. The free snapshot gives you three verified findings from your own site. The Website Revenue Audit (£295) gives you the full evidenced picture. And if you would rather not do the fixing yourself, the Repair Sprint (£795) implements the fixes for you. Or simply ask us directly.