Find Content Opportunities in Search Console with AI

A page with falling clicks does not always need a rewrite. Search demand may have declined, its query mix may have changed, or a different search result layout may be attracting attention. Updating the copy before checking those explanations can waste the opportunity you actually have.

Google Search Console helps you find content that deserves investigation. It shows how your site appears and receives clicks in Google Search, within the report's definitions and data limits. The useful output is a short list of evidence-backed edits, not a spreadsheet of every query with an average position between eight and twenty.

Give Claude or ChatGPT a consistent Search Console export, then use RankParse to inspect the public pages behind the opportunities. The agent can group queries and draft changes; your job is to verify the comparisons and choose edits that solve a real reader problem.

Start with an export in Claude or ChatGPT

Connect RankParse using its MCP setup guide, and upload your Search Console exports to the conversation. The default secure OAuth connector does not expose Search Console tools. In this workflow, account performance comes from your uploaded files; RankParse adds public-page context through tools such as get_page_seo.

Export the page comparison and the query tables for the few pages you want to investigate. Label each file with its property, date range, search type, and filters. Separate page and query exports are not automatically a query-to-page mapping: export queries with the relevant page filter applied, or supply a dataset that actually contains both dimensions.

Give the agent this assignment after choosing the comparison settings below:

Analyze my uploaded Search Console exports for [business objective].
First list each file's date range, filters, dimensions, and missing data.
Do not join queries to pages unless the files support that relationship.
Group queries by reader task and select at most five page opportunities.

For those URLs, use RankParse page SEO checks for current title, headings,
and canonical context. Use supplied page text or available browsing to
assess the actual answer. Separate historical performance from current
page observations. Never invent Search Console data from public tools.

Return: page, query group, observed movement, alternative explanations,
reader problem, proposed edit, supporting source, confidence, and review
plan. Recompute CTR from clicks and impressions, not an average of row
percentages. Do not claim causation or guaranteed additional traffic.

Check the assistant's first calculation against the original export. Confirm that it preserved filters and did not silently treat missing query rows as zeros. Public-page checks show the current page, which may differ from the version that generated the historical performance.

An advanced alternative is RankParse's API-key MCP route in a compatible developer client, with Search Console connected separately. That route offers account tools; it is distinct from the standard Claude or ChatGPT web connector. An uploaded export is sufficient for the workflow here.

Give the assistant a packet it can reconcile

Start with three to five pages serving one business objective, such as helping finance teams create an expense policy. Upload the page-level comparison, one page-filtered query export per candidate, current page text, and a brief change log. The connector will supply the current page checks; the files supply the historical performance and editorial context.

A page filter recorded in a filename or manifest is useful provenance even when the query CSV has no URL column. Make that relationship explicit. Conversely, a sitewide query CSV cannot be attached to every candidate page merely because its words resemble their titles.

First pass: reconcile the uploads before using credits

Audit the packet before interpreting opportunities. List each file's
scope and supported joins. Recalculate one page's CTR for each period
from clicks divided by impressions, and show the arithmetic. Check for
zero denominators, missing rows and overlapping query groups. Report
unmapped queries separately. Do not call public tools until you can name
the candidate page URLs and the question each check should answer.

Use this pass to catch bad comparisons early. If the periods use different device filters, re-export them; a confident narrative cannot repair incompatible inputs. If query rows do not sum to page totals, preserve both totals and the reporting limitation instead of scaling rows to make them match. For a low-volume site, it may be reasonable to make an editorial improvement without a search-performance claim.

Second pass: build a current-page evidence card in RankParse

For each shortlisted URL, ask Claude or ChatGPT to use get_page_seo and save status_code, final_url, canonical.url, robots.raw, title.text, description.text, headings.h1, and headings.h2. These page-check fields answer distinct questions: did the intended page respond, does it declare the expected URL, and does its fetched promise resemble the task in the query group?

For the shortlisted URLs, run get_page_seo and create a current-page
card with the requested fields and collection time. Compare its title
and headings with the uploaded query groups and captured page text.
Separate a response/canonical discrepancy from an editorial opportunity.
Quote the source supporting any claimed missing answer. The first ten H2s
are not the whole article; do not infer absence from that list alone.

Read the result against the export URL. If final_url now differs, consult the migration log before assigning a rewrite. If the canonical points elsewhere, investigate the relationship and obtain Google's reported canonical separately when relevant. The public check cannot tell you how Google attributed historical performance.

A failed fetch belongs in an “evidence needed” queue. A 200 response with a challenge title also needs investigation: the tool may have parsed an access screen. When the fetched HTML lacks client-rendered content, supply the browser-visible copy. Do not label either case a confirmed content gap.

Third pass: challenge the recommendation before briefing an editor

Review each proposed edit against the export and RankParse evidence card.
Give the strongest alternative explanation for the movement. Name the
one additional observation most likely to change the decision. Classify
it as ready to edit, technical investigation, or insufficient evidence.
For ready rows, provide a short brief: existing promise, reader problem,
exact section to change, source material required, owner, acceptance
check, and later comparison settings. Do not publish changes.

This is where the two evidence sources become useful together. Historical impressions identify a reader task worth investigating; the current-page card and supplied copy tell the editor whether the page still makes and fulfills that promise. Neither alone establishes why clicks changed.

With a small team, work one ready brief at a time. For a larger site, group the evidence cards by template and owner. If many cards show the same unexpected canonical or title pattern, pause individual copy briefs and investigate the shared template. Keep exceptions visible instead of forcing every page into the majority explanation.

Set up a comparison you can explain

Choose one search type and a date range with enough activity to interpret. For a regularly visited site, two complete four-week periods can be a starting point. Low-volume pages may need longer periods; seasonal businesses may need a year-over-year comparison as well.

Keep country, device, search appearance, and other relevant filters consistent. Do not compare an all-device period with a mobile-only period or a partly complete week with a complete one. Treat very recent or preliminary data cautiously.

Save the comparison settings with the export. Record deployments, substantial content changes, promotions, and events that might affect demand. You do not need a perfect experiment to make a decision, but you do need to know what you compared.

Keep the business objective beside the comparison settings. Recovering visits to a service category and finding unanswered guide questions require different opportunity queues.

For a broader view of those goals, connect this exercise to SEO reporting that drives decisions.

Know what the numbers can and cannot tell you

Clicks, impressions, click-through rate, and average position describe different parts of search visibility. They do not measure the same outcome as analytics sessions, leads, or revenue.

Google's Performance report documentation explains the metrics and aggregation rules. Average position uses the topmost result associated with the relevant property or page for an impression, depending on aggregation. It can shift because the queries producing impressions changed, even when placement for an established query did not.

Averages also hide variation between countries and devices. A page could improve for its core query while gaining impressions for many new queries in lower positions. Its overall average position might worsen while its reach expands.

Search Console omits some queries for privacy and limits the rows available through its reporting interfaces. Query rows therefore do not represent a complete list of every search, and their sum can differ from a chart total. Filtering can change what data is included. Treat an export as the observed dataset, not a complete census.

Performance is generally attributed to the canonical URL Google selects. If a page seems to have lost data, investigate canonical changes or URL migrations before assuming the content lost its audience. Page and property aggregation can also produce different totals; preserve the same scope when comparing.

Build query groups around a reader's task

Open the Pages tab and select a page with a meaningful change or a clear business role. Then inspect the queries associated with that page. Group related queries by the question or task they express.

A guide about expense policies might appear for several distinct groups:

• Examples and templates: readers want something they can adapt.
• Approval rules: readers want to define who authorizes spending.
• Reimbursement timing: readers want a process or expectation.
• Software queries: readers are evaluating tools.

These groups imply different content decisions. Adding a template could serve the first group. Expanding approval examples might serve the second. A tool comparison could deserve a separate page if it represents a genuinely different intent.

Start manually with a small number of pages. Use filters or regular expressions once you understand the language. Review grouped queries so that a shared word does not merge unrelated needs. Keep branded and nonbranded searches distinguishable when they tell different stories.

Look for four patterns worth investigating

Existing visibility, incomplete answer

A page receives relevant impressions for a query group but only partly answers the question. Read the page as a visitor would. Is the answer buried, missing an example, out of date, or too general to use?

Check the live results for representative queries in the target market. Results vary by location, device, and time, so a manual check is contextual evidence rather than a universal ranking measurement. Note whether prominent results are guides, product pages, templates, videos, or another format.

If your page serves the right intent, improve the missing section. If it serves a different intent, inserting more keywords is unlikely to solve the mismatch. Your competitor analysis should explain what readers can accomplish with the competing results and what your page can add.

Stable visibility, weaker click response

A query group retains similar impressions and observed placement but receives fewer clicks. Investigate whether the title accurately describes the page, whether the promise is specific, and whether the search results have changed.

Compare similar devices, countries, and search appearances before labeling the problem “low CTR.” There is no universal CTR target for a position. A branded navigational search behaves differently from an informational question answered directly in the results.

An improved title can make a result clearer, but Google may generate a different title link or snippet. Treat title and description edits as a test of a specific promise, not a guaranteed increase in clicks.

A useful page losing relevant queries

A page loses clicks and impressions for the same established query group. Check whether placement also changed and whether the pattern appears across related pages.

Inspect technical access, indexing, and canonical behavior before commissioning a rewrite. The technical SEO audit checklist helps distinguish a content issue from a template or indexing problem.

If the page remains technically sound, inspect factual freshness, missing requirements, changing intent, and competing answers. Search Console alone cannot establish the cause. It identifies where to investigate and which questions to ask.

Several pages appearing for one task

Multiple pages receiving impressions for similar queries can be healthy. A guide and a product page may satisfy different needs. It becomes worth investigating when near-duplicate pages compete for the same role or the site repeatedly sends readers to a less useful destination.

Inspect the query's Pages view and read the candidates. Do not call it cannibalization solely because two URLs appear. If their purposes overlap substantially, consider merging useful material into the strongest suitable destination and redirecting only where the replacement genuinely serves the old page's audience.

A worked example: choose an edit, not a ranking target

The following numbers are illustrative, not measured customer results. A team reviews an expense-policy guide across two complete, comparable 28-day periods. Search type, country, and device filters are unchanged.

The table summarizes observed query rows, so it need not equal the page-level report total. These numbers identify questions; they do not establish causes.

For the template group, the assistant calculates earlier CTR as 160 / 4,000 = 4% and later CTR as 126 / 4,200 = 3%. The decline is one percentage point, or 25% relative to the earlier CTR. That calculation describes the supplied rows; it does not explain the decline.

In this illustrative case, the RankParse evidence card contains status_code: 200, an unchanged final_url, the expected canonical.url, and title.text: "Expense policy template and approval guide". Its H2 list includes “Approval rules” and “Reimbursement timing.” This narrows the next check to the promise made by the title, but the heading list cannot prove a template is absent.

The editor reviews the supplied full copy and confirms that it contains only explanatory prose, with no reusable template. The team also compares average position across matching query, device, and country groups and reviews live results. It now has a defensible reader problem even though the cause of the CTR movement remains uncertain: the title promises something the page does not deliver. The brief is to add a usable, clearly scoped template and make the title match the resource.

For approval rules, the team sees growing visibility and a stable click rate. There is no immediate reason to rewrite the section. It checks factual completeness and leaves it alone if readers can use the answer.

For reimbursement timing, unchanged CTR alongside fewer impressions suggests examining demand and placement before changing the snippet. The team checks the same period in the previous year and comparable pages. If evidence remains inconclusive, it records uncertainty instead of inventing a content failure.

Ask the agent to produce those decisions as separate rows, citing the export group and the page evidence behind each one. Review the proposed template yourself before publishing. This produces one concrete improvement and two investigations. It avoids a blanket rewrite that would make later interpretation harder.

Keep the opportunity queue small

For each candidate, record the page, query group, observed change, suspected reader problem, proposed edit, and evidence still needed. Add a business relevance assessment and the estimated work involved.

Prefer a well-supported improvement to an important page over a speculative fix to a large but irrelevant impression count. Do not multiply all impressions by an assumed “ideal CTR” and present the result as a traffic forecast. That calculation can be a scenario, but its assumptions need to remain explicit.

A useful action note looks like this:

Add an editable expense-policy template to the existing guide because template queries already produce relevant impressions and the current page fails that promise. Keep the main URL. Check the template for legal and operational scope, update the title to describe the resource accurately, and record the publication date. Review the same query group after a comparable observation period.

When an edit creates a genuinely useful destination, connect it from relevant pages using an internal linking strategy. Avoid treating new content and existing site navigation as separate projects.

Review the result without forcing a success story

After the editor publishes, repeat get_page_seo for the edited URL. Compare the returned title, description, canonical, response and headings with the accepted brief. Open the page to check the actual template or new answer; metadata alone cannot verify it. Retain before-and-after cards alongside the publication date. The read-only connector supports these checks but does not apply the edit.

Later, upload a new Search Console export with the saved filters and ask the agent to compare the same page and query groups. Keep newly observed queries separate from the established group so a changing mix does not silently redefine the test. Look at clicks and impressions together, then examine CTR and position in context.

Allow enough time and activity for a meaningful comparison. There is no fixed number of days that makes every content test reliable. Sparse data, seasonal demand, and Google's recrawling schedule can require a longer observation period.

Record whether the evidence supports keeping the edit, extending it, or investigating another explanation. A better template might improve visitor usefulness without producing an immediate search lift. A traffic lift might coincide with unrelated demand growth. Report those distinctions plainly.

Choose one ready brief, save its evidence card, and schedule the next comparable export before expanding the queue.