Traffic arrives, but the page still fails: diagnose intent mismatch

Define the failure after the click

A page can receive impressions and visits while failing the job that attracted those visits. Begin with an observable failure: readers return immediately to search, ask a question the page claims to answer, choose an unrelated CTA, or reach the bottom without locating required evidence. Avoid naming ‘wrong intent’ as the cause. Preserve the query-page sample, result snippet, rendered page, event definitions, support language, and recent edits. The diagnostic artifact is a promise-break log with columns for expected job, promise location, observed obstacle, competing explanation, discriminating check, and repair owner. Google’s people-first guidance makes audience usefulness a relevant standard, while usability evidence helps test whether people can complete realistic tasks. Neither turns a bounce metric into a diagnosis.

Evidence: Google Search Central; GOV.UK Service Manual

Separate dashboard signals from reader failure

The log should contain failures as well as successful sessions. Averages can hide a small but important audience whose safety or accessibility needs are being missed.

Evidence: Google Search Central; Google Search Central

Hypothesis one: the result promise attracts another job

Compare the title and description as displayed with the first answer on the page. A broad superlative, ‘complete guide,’ or ‘best’ claim may recruit comparison shoppers to a definition article. Conversely, a technical label may hide a practical worksheet. Sample query language by decision, then ask task participants what they expect before opening the result. This hypothesis gains support when expectations consistently differ from the page’s actual completion state and the page performs coherently once the intended job is explained. It weakens when participants predict the correct job but still cannot complete it. Do not copy competing result language without checking whether your evidence can fulfil the copied promise. Google’s starter guidance favors descriptive, useful presentation; it is not permission to overstate breadth for a click.

A repair candidate is a narrower result promise or a genuinely expanded answer, not superficial keyword substitution.

Evidence: Google Search Central; World Wide Web Consortium

Hypothesis two: the answer is hidden behind context

Inspect the first screen, heading outline, and the route from problem recognition to usable answer. Readers may understand the promise yet abandon because the page starts with generic history, advertising, or an unexplained table. Run a find-the-answer task and note the first heading selected, scroll reversals, and verbal confusion. W3C’s headings guidance supports labels that describe topic or purpose and help people locate content. The hypothesis is supported when the necessary answer exists, can be used after it is revealed, but is repeatedly missed or delayed. It is weakened when the located answer lacks evidence or cannot settle the task. The repair may move a bounded answer earlier, rename a heading, or split prerequisites into a clear branch. It should not remove a safety qualification merely to reduce scroll depth.

Record the exact barrier in the promise-break log; ‘poor engagement’ is too vague to guide an accessible content change.

Evidence: Google Search Central; World Wide Web Consortium

Hypothesis three: the page asserts more than it proves

A visitor may reach the intended answer and reject it because the evidence does not support the confidence, scope, or commercial recommendation. Inventory material claims, their sources, relevant dates, and visible limitations. Separate official definitions from independent analysis and illustrative examples. For affiliate pages, confirm that product comparisons use disclosed criteria and that the conclusion does not exceed the collected evidence. This hypothesis gains weight when participants can restate the answer but ask for missing proof, exceptions, or applicability. It weakens when the evidence is found and understood while the task still stalls at navigation or action. More citations are not automatically better; the source must address the actual claim. A generic trust badge cannot substitute for provenance.

The discriminating repair is to narrow the claim, strengthen the evidence, or mark the uncertainty—not to intensify the CTA.

Evidence: Google Search Central; GOV.UK Service Manual

Hypothesis four: the final action jumps a decision stage

Trace what the reader knows immediately before each invitation. An article may explain a concept responsibly and then present a purchase button as though a product decision has already been made. That stage jump can create exits, accidental clicks, or distrust even when the body is aligned. Test whether the CTA describes a logical continuation, discloses commercial relevance, and lets the reader predict the destination. W3C’s heading and label principles help evaluate clarity, while task observation reveals whether people interpret the action as intended. Support for this hypothesis comes when the answer is accepted but the next step is described as premature or unrelated. It weakens when the CTA is ignored because an earlier promise remains unsatisfied. The repair can offer a comparison worksheet, criteria guide, or optional product reference instead of forced urgency.

Do not interpret a higher click rate by itself as proof that the new action is more useful or ethically better.

Evidence: Google Search Central; GOV.UK Service Manual; World Wide Web Consortium

Close the log with a falsifiable cause

For one failing page, select the branch supported by the strongest direct observation and write what would disconfirm it. Make one reversible change at the earliest broken promise, then repeat the same representative task and technical checks. The next action is to complete a single promise-break row and schedule its evidence review; multiple simultaneous rewrites stay in backlog. Limits are unavoidable: analytics cannot expose motive, small task samples may miss rare needs, result snippets can be rewritten by search systems, and seasonal traffic can mimic an alignment change. If two causes remain plausible, preserve both and gather the cheaper discriminating evidence before editing. A diagnosis is complete when the symptom, cause, repair, contradiction, and observation window can be reconstructed. It is not complete because a dashboard moved in the preferred direction.