What a fidelity score measures

Ask for London-based technology decision makers and every channel will sell you something. What none of them will tell you unprompted is how far the thing they sold you is from the thing you asked for. On connected TV it is a set of consent-chained business intent segments delivered against streaming households, which is a good proxy. On a billboard it is a rail station at commuter hours, which is a statement about a place rather than about a person. Both are legitimate buys. They are not the same buy, and no invoice tells you which one you got.

A fidelity score is the answer to that single question, computed at the moment the audience is compiled rather than asserted afterwards. AdBuyMCP's compilers return one with every channel plan they produce, and every band is published, with the condition behind it, in the fidelity dataset.

The number measures the targeting and nothing else. It says nothing about the inventory, the viewability or the attention, and it makes no claim about reach or outcomes. It is also a property of the channel's vocabulary, so a competitor buying the same channel faces the same ceiling whether or not they publish it.

Each score carries a grade naming what is really being addressed. The grades run from closest to the individual to furthest from it, and they are the part of the model that survives translation into a slide.

These are all the bands the catalogue's compilers return, each channel's best band first, with what a brief has to supply to reach each one. Four channels return a single band whatever the brief says. The other three move with it.

Each channel has a compiler, and a compiler is a pure function: persona plus registry in, targeting specification plus fidelity object plus lawful-basis manifest out. Same input, same output, every time, which is what makes a plan auditable and a disagreement about it resolvable.

The compiler matches your persona's own terms against the keywords on each catalogue entry and takes the top few detailed matches: four attributes on addressable TV, three packs on cinema, four venue types and three audience segments on out-of-home. Match confidence is not decoration: on connected TV a person-level grade requires at least one match at or above the strong-match threshold, and a set of weak matches produces a household grade and a flag for human review rather than a promotion.

The important structural rule is what the language model is not allowed to do. It extracts semantics from your brief (geography, seniority, sector, intent topics, exclusions) and it never emits a vendor segment identifier. Deterministic registry tables do that. A model can hallucinate a segment id that looks exactly like a real one; a lookup table cannot return an entry that is not in it, and that is the difference between a plan you can audit and a plan you have to trust.

Three things move a score, and the compiler names each one in its rationale or in a warning rather than absorbing it.

The first is the quality of the brief, on the three channels where it matters. Paid search compiles to 82 from buyer-authored intent topics, 68 from interests, and 45 from a persona label with no product context. At 45 the rationale says exactly what is missing, which is product and offer context before the keyword plan is activation-ready. On the four fixed channels no amount of audience detail changes the score at all, and pretending otherwise would waste a buyer's afternoon.

The second is geography we cannot map. Naming a city with no entry in the geography table costs fifteen points, floored so a score cannot fall below ten, and the rationale names the city that caused it. The alternative, silently dropping an unmapped city and reporting the same score, is the failure mode that makes a plan look complete and deliver nationally.

The third is an exclusion the channel cannot express. Where your brief excludes something and the channel has no native negative targeting, the compiler drops the overlapping segments and raises a warning naming the exclusion and what it dropped, rather than quietly buying an audience you asked it not to.

A fidelity score answers how close the targeting got. A lawful-basis manifest answers what made it lawful, and the two are produced by the same compilation pass because they are the same question asked from two directions.

Every segment that enters a plan goes through a policy gate first. A consent-chained source is allowed with a condition attached: the supply chain must be consent-attested and the vendor attestation verified on connection. A legitimate-interest source is allowed with the condition that a legitimate interests assessment, the test the ICO describes for that basis, is on file for it. Aggregate and buyer-authored inputs (cinema audience packs, out-of-home venue types, postcode polygons, your own search terms) carry no personal data and are allowed outright, which is why cinema's manifest is the shortest in the catalogue.

The manifest also carries a baseline entry for the channel itself and for the geographic scope, so the answer to "where did this audience come from?" is a document attached to the plan rather than a conversation weeks later. It is a record of what the compiler acted on. It is not a legal opinion and it does not replace your own assessment.

Cinema scores 40 because it targets by venue, and it can still be the right buy: a full screen with the lights down has an attention profile nothing else on the list matches. The score measures only the targeting, so it tells you what claim a channel can support afterwards and leaves the question of whether to buy it to you.

The same holds for out-of-home. A billboard cannot know who is walking past, so its score describes a place and a time, and a buyer who wants to reach a business district at commuting hours is buying exactly what the score says.

Read the grade before the score. Scores of 68 at content grade and 65 at household grade are close as numbers and different as buys, and the grade tells you which claim the campaign can support afterwards.

Then check whether the band is fixed, as it is on four of the catalogue's channels. If it is, your effort is better spent on the brief for the channels where it moves (search, CTV and paid social) than on trying to improve a number that is a constant.

Then read the rationale and the warnings, which is where the compiler explains itself. An unmapped city, a dropped exclusion, a weak segment match flagged for review: each of those is a specific thing you can fix by editing the brief, and each of them is invisible on a plan that reports a single blended figure.

Finally, use the score to size the argument rather than to rank the channels. The right question is not "which channel scores highest?" but "is this score good enough for the money I am about to put behind it?" Only you can answer that.