# Quality Score diagnostics — research and method Last reviewed 2026-09-29. ## The governing refusal Google documents Quality Score as a diagnostic of predicted relevance, and states that it does not predict performance. Every design decision here follows from taking that seriously rather than paying lip service to it. If Quality Score does not contain information about cost, conversions or return, then a tool that derives any of those from it is manufacturing a number. So this tool produces no predicted cost per click, no predicted conversion rate, no predicted return, and no Ad Rank figure. The refusal is not a disclaimer at the bottom of the page; it occupies the third preview statistic, permanently reading "Not predicted", so the gap is visible exactly when someone is looking for a number to fill it. The same rule is applied to the score itself. A blank, negative or above-10 value is reported as unscored rather than clamped into a band. A band that looks authoritative while being wrong is worse than no band at all, and a user who types 11 is either transcribing something else or working from a different scale entirely. ## Asserted as documented platform behaviour **The four bands.** Above 7 is Excellent, 4 to 7 is Good, 1 to 4 is Needs improvement, below 1 is Poor. The bottom of the "4 to 7" band is inclusive, so a score of exactly 4 is Good. There is no separate band for exactly 10, and the page says so, because "get to 10" is an expectation the platform never created. **The three factors.** Ad relevance, expected click-through rate, and landing page experience. All three are estimates from historical signals, and none is a measurement of a conversion. **Per-keyword and per-placement reporting.** Ad relevance and expected click-through rate are assessed per keyword, not per ad group, so a strong ad elsewhere in the group does not lift a weak one. In Display and Video, Quality Score is calculated per placement group, and only for placements where it applies. **Two diagnostic scales.** Search reports the three factors as above or below average. Display and Video report them on a five-step scale: very low, low, below average, above average, high. These are different instruments reading different questions, and the tool asks which one applies before it explains anything. A Display score of 8 feeling better than a Search score of 8 is pure category error, and the page says so. **Smart Bidding requires a conversion history** in the campaign before it will run, which is a requirement separate from holding a Quality Score. **Quality Score is not the page experience score.** They overlap and are routinely conflated. Page experience concerns speed, responsiveness and visual stability; Quality Score's landing page experience factor concerns relevance to the ad and the query, with load time as one part. Google recommends Core Web Vitals for measuring page experience, and Quality Score is not a Core Web Vitals score. ## Asserted as our own judgement, and labelled as such These are opinions about where effort pays off. They are not facts about the platform, and the page presents them as recommendations rather than rules. - **The order of the fix list.** Ad relevance first, because fixing copy is cheap and also moves expected click-through rate. Landing page last, because page speed is genuinely slow to change and is frequently blamed for relevance problems that ad structure caused. - **That improving a score usually costs more per click.** A better-matched ad tends to be clicked more in a competitive auction, and that is what the improvement is worth. This is the commercially important claim on the page, and it is the reason "fix every low score" is bad advice. - **That a profitable keyword at a low score should often be left alone.** This is the single most counter-intuitive point made here, and the one the product exists to make: on brand and navigational terms a low score is close to normal, and a score improvement on a keyword that already pays back is often a downgrade. The page separates the two explicitly. The band edges and the factors are labelled "documented"; the ordering and the cost note are labelled "judgement, not a platform rule". ## Not asserted anywhere - The formula or weights combining the three factors. Not published, not invented here. - Any Ad Rank figure. Ad Rank is commonly described as bids multiplied by Quality Score, expected click-through rate and other factors, and the page labels that as a teaching model rather than a formula, because the real calculation and the weight of each part are not published. - Any predicted cost per click, conversion rate or return. See the governing refusal above. - Any target score. The page says an account full of brand terms does not need one. - Time to improvement, expressed as a number. Quality Score is calculated on historical signals and moves only after a change has gathered enough impressions to measure, which is described qualitatively as weeks rather than given a false precision. ## Known limits - The tool takes the three diagnostics as the operator reports them. It cannot read the account, so it cannot detect a mis-transcribed rating. - A diagnostic of "not sure" is treated as unreported, which is the right default and still leaves open the possibility that the operator simply has not looked. - The causes and fixes listed per factor are the commonly documented ones, not an exhaustive set of every way a factor can be rated down. They are written to cover the frequent cases, and the page does not present them as complete. - Band definitions have not changed in Google's documentation for a long time, but platform behaviour does change. The page invites users to trust their account over this page and to report a discrepancy. ## Verification `tests/run.mjs` checks all four band edges and both boundaries of each band, the refusal across several states, the absence of any currency figure or Ad Rank claim in the rendered results and in the exported CSV, refusal of out-of-range scores, both diagnostic scales and the reset-on-scale-change behaviour, the distinct treatment of an unreported diagnostic, the absence of manufactured work on a clean score, the fix ordering, and a real CSV export.