Start with a verified problem that prevents a customer from contacting you or prevents an important public page from being accessed. Then improve the pages that already attract relevant searches. Cosmetic warnings and speculative AI scores can wait.
That is our decision rule, not a promise that every site needs the same work. A report with 200 warnings does not mean you need 200 separate projects. Here is how to turn it into a short, testable plan.
1. Can a real inquiry reach your business?
Open the page on a phone and follow the contact path. Is the offer understandable? Can you use the form? Does the business receive the information? A button that says “Sent” is only one part of that check.
Test with a clearly labeled internal submission, agreed with whoever receives the messages. Check the saved record and delivery destination. Record it as a test so it does not become a claimed customer. Do not put personal contact details in Analytics.
If the report says zero leads, ask whether the source contains the full selected period. A limited list of recent submissions can produce a misleading zero. Separate saved forms, team tests, spam and inquiries about your actual services. A phone-button click is not a confirmed call.
Fix first when: A reproducible failure blocks a valid inquiry. Keep evidence of the failure and the successful retest. If only the reporting is wrong, repair the count before making claims about demand.
2. Can search engines access the pages that matter?
Ask the developer to inspect the exact service URL named in the report. Check the response, access restrictions, indexing instructions and whether the main content is available. Do not change every setting because one audit tool paints it red.
Google’s minimum Search requirements include accessible content, a successful HTTP response and indexable content. Meeting them does not guarantee indexing. For its generative AI features, Google also describes indexing and snippet eligibility as requirements.
Google’s technical requirements
Google’s guidance for generative AI search
Fix first when: An intended public service page is unintentionally blocked. Preserve access restrictions on private areas; visibility is not a reason to disable security across your site.
3. Match the fix to the search problem
Look at the page and the queries that actually show it in Search Console. Compare equal, complete periods and keep the country, device and language filters consistent. Impressions show appearances, not people who read your page.
Relevant searches, reasonable visibility, few clicks: Compare the search result with the page. A title should describe the problem it solves or the service it offers. Change a vague title when you can explain the mismatch; do not add claims the page cannot support.
Relevant searches, low positions: A new description alone is not a complete plan. Review whether the page answers the question well, contains useful evidence and is linked from relevant pages. Low visibility can limit clicks even when the wording is clear.
Visits, but few useful inquiries: Review the offer, trust evidence and next step. Check the complete received-form history before concluding that nobody contacted you.
Google can generate a title link from several page signals. Your edited title is an input, not a guarantee of the exact text shown in search.
4. Make AEO findings specific enough to act on
“Your AI visibility is low” is not a repair instruction. Ask for the tested question, tool, date, language, answer and source links. Was your business mentioned, recommended or linked? Those are different observations.
If an answer misunderstands your service area, inspect the page it cites. Correct a demonstrably outdated statement and add a documented example when you have permission to show it. If the answer cites another source, record that limit rather than claiming your page caused the result.
Keep a stable set of neutral buying questions and repeat the checks in fresh sessions. One answer changing after an edit does not prove the edit caused it. Do not buy a guaranteed recommendation or create several near-identical pages for variations of one question.
5. Give each proposed fix an owner and a finish line
Copy these six fields into your task list. If the report cannot fill them in, the next action is investigation rather than a large implementation.
Page and problem: The exact URL and a plain-language description of what fails.
Evidence: A reproducible check, dated screenshot or report with its filters. An automated warning is a lead to investigate.
Business consequence: Which customer action or discovery path is affected? Prioritize important pages and shared components over isolated low-value warnings.
Action and owner: One bounded change, who makes it, the expected effort and any dependency.
Acceptance check: What must work after the change? Include mobile, the other language when relevant, and any form involved.
Outcome review: Record the baseline and review comparable periods after release. Fix confirmed breakages immediately; allow traffic outcomes enough observations to be meaningful.
An example: three findings, one sensible order
This is a hypothetical example, not a client case or a forecast. A service website has a contact form that fails on mobile, a service page with relevant impressions but low positions, and twenty decorative images without descriptive text.
First, repair and retest the mobile form. Second, inspect the service page’s query intent and missing evidence before choosing its improvement. Third, assess the image warnings individually: meaningful images need appropriate alternatives, while purely decorative images may appropriately use empty alternative text.
The number of warnings did not decide the order. The verified customer obstacle did. A different site with an inaccessible homepage could need a different first action.
What should you measure after the fixes?
Keep a release log: URL, date, change and expected outcome. Track search impressions and clicks separately from sessions. Then review received forms and actual service inquiries, with team tests and spam clearly identified.
A practical starting point is equal complete 7-day and 28-day windows, provided the sample supports interpretation. Record other changes, campaigns or seasonal effects. Publishing successfully proves the change is online; it does not prove more customers came from it.
If speed is the demonstrated obstacle, review our speed service
Questions before you approve the work
Should I fix every warning in an audit?
No. Verify each important finding and group shared causes. Prioritize failures that affect customers or important public pages, then improvements supported by relevant search and behavior data.
Should I rewrite all my titles if impressions rise but clicks do not?
No. Review queries, positions, devices and intent first. Change titles where there is a clear mismatch, then measure comparable periods without claiming that the title controls every click.
Does a better audit score guarantee AI recommendations?
No. A score is a tool’s assessment. Check the underlying evidence, actual answers and business inquiries separately.
