Quality Score: Fix the Inputs, Not the Number

Quality Score is a diagnostic readout wearing a KPI costume. The one-to-ten number is a summary of three rated components — expected clickthrough rate, ad relevance, landing page experience — and the number itself is not what you fix. The components are. Pull all three for every keyword in one prompt, group keywords by their weakest part, and the vanity metric turns into a sorted fix list. That reframe is the whole method; the rest is execution.

What Quality Score is — and why chasing tens is a trap

Quality Score rates each keyword one to ten from three components, each graded Below average, Average or Above average — the structure is laid out in Google's documentation. The underlying quality signals feed Ad Rank, so they shape position and cost per click. That much makes the components worth real attention.

The trap is treating the visible number as the goal. Tens live where relevance is free — brand terms, navigational queries. Competitive commercial keywords with perfectly healthy components routinely sit at six or seven, and "improving" them past that point produces algorithm-pleasing copy and cowardly keyword lists. The productive target is different: no keyword failing a component you could fix. Aim there and the number follows as a side effect — which is the correct direction of causation.

Pulling QS components account-wide in one prompt

The interface shows components keyword by keyword, which is why nobody reviews them. Connected, it is one ask:

For account 123-456-7890, pull every enabled keyword with Quality
Score and all three components: expected CTR, ad relevance, landing
page experience. Group keywords by their weakest component, rank
groups by total spend, and within each group show the worst
offenders. Note keywords with no score separately — that is data
absence, not failure. Do not recommend fixes yet; show me the map
first.

The output is the diagnosis: how much spend sits on keywords failing each component. Spend-weighting is what makes it actionable — a Below-average component on a keyword spending heavily outranks a dozen fours on keywords spending nothing. Map first, then fix by group, because the three components fail for different reasons and reward different work.

Snapshot the map monthly. Google exposes historical quality columns per keyword, but the grouped, spend-weighted account view exists only when you build it — and a dated series of maps is what lets you say the ad-relevance work in March actually moved components by May, rather than remembering it fondly.

Fixing expected CTR: relevance and match tightening

Below-average expected CTR means Google predicts searchers will not click — historically, ads on this keyword's matched queries have not earned the click. Two working levers:

• Tighten what the keyword matches. Loose matching pulls the keyword into queries where your ad is a weak answer, and those auctions drag the prediction. Read the matched search terms, add negatives for the off-intent themes, and consider stricter match types where drift persists.
• Earn the click harder. Headlines that echo the query family, a concrete offer, specifics a competitor cannot copy. This overlaps ad relevance work below, but the emphasis differs: expected CTR is about desirability, not just topical overlap.

Structural honesty helps here too: a keyword stranded in an ad group whose ads serve a different intent will never predict well — move it to where the ad actually answers it. A model with the whole account loaded finds these strandings in one pass: ask for keywords whose component failure coincides with an obvious ad-group mismatch, and the relocation candidates fall out as a list.

Fixing ad relevance: RSA coverage per ad group

Below-average ad relevance is the mechanical failure: the ad does not talk about what the keyword means. The fix is correspondingly mechanical — close the gap between the keyword's language and the ad's headlines. In practice that means keyword coverage in the RSA slate: headlines that include the keyword's phrase family naturally, alongside the proof and offer lines.

The efficient route is the RSA drafting workflow: have the model read the ad group's keywords and converting search terms, then draft headlines in that language. Where one ad group's keywords have drifted into multiple intents — the usual root cause — split the group so each RSA can be about one thing. Ad relevance is the component most fully under your control, which makes it the fastest win on the map.

Fixing landing page experience without a rebuild

Below-average landing page experience rarely requires the redesign everyone fears. The rateable problems are usually narrower: the page does not mention what the keyword promises, arrives slowly, or fights the visitor on mobile. The corresponding fixes are copy and routing before they are engineering — make the page state the offer the ad made, in the searcher's vocabulary, above the fold; and route keywords to the most specific page that answers them rather than the homepage.

Ask the model for the mismatch audit: "For keywords failing landing page experience, compare each keyword's intent against its final URL's actual content and list the gaps." The output is a routing-and-copy worklist, most of it doable without a developer. Genuine speed and mobile failures do need engineering — but find out whether relevance was the problem first, because it usually was.

When to just retire the keyword instead

Some low-QS keywords are not fix candidates — they are mistakes with tenure. The tells: intent you cannot actually serve, a component failure that survived two rounds of fixes, spend without conversions across a long window, or a keyword whose queries a better keyword already covers.

Retirement is judgement, so run it as a review: "List low-QS keywords where fixes have not moved components, with spend and conversions over 90 days — recommend keep, fix again, or retire, with reasons." Pause the retirees through the connector with one approval — pausing preserves history and is fully reversible, and permanent removal stays a deliberate human act in the interface, which is exactly the boundary a well-designed AI setup enforces anyway. The review is also where keyword lists stop growing monotonically: accounts accrete keywords for years because removal requires a decision, and now the decision has a standing slot.

A quarter of this — one map, three fix passes, one retirement review — typically clears the avoidable component failures. After that, Quality Score goes back to being what it always was: a gauge you glance at, reporting on inputs you already manage. That is the healthiest possible relationship with a score — attention proportional to what it can teach, and none spared for the number itself.