After a site migration: automated Search Console reporting under SEO control

The website of Havana Club had just been rebuilt. Its structure had changed. For the brand, the question was no longer simply whether the migration was finished. The team needed to see, week by week, whether organic visibility was recovering - and which changes deserved intervention.

There was no manual reporting process to replace. Monitoring was designed with the post-migration need, before a recurring collection and formatting routine had time to form.

IZZY built a weekly workflow that retrieves Google Search Console data, compares the latest snapshot with every previous one, builds recovery trends, creates charts and generates a report according to the reporting brief. An LLM layer carries out the analysis and drafting automatically.

Before the document reaches the client, an SEO specialist reads it, simplifies it when necessary and approves it. In most cases, that review takes 15 to 25 minutes.

Report production is automatic. Delivery is not.

That distinction matters: a report produced without clicks is not a diagnosis produced without judgement.

A drop in clicks can reflect lower demand, a position change, a page problem, a different result-page presentation or several factors at once. Search Console observes some of the consequences inside Google Search. On its own, it cannot reveal the whole cause, what happened after the click or the commercial outcome.

The useful objective is not to ask AI to be right on its own. It is to make observation repeatable, expose the limits and reserve human time for the changes that deserve investigation.

The answer in 60 seconds

• The workflow was created to monitor a restructured website after migration; it did not replace an established manual report.
• It runs weekly and compares the current snapshot with every previous snapshot to build a recovery trajectory.
• The reporting brief defines the rules; collection, charts, analysis and drafting then run automatically with an LLM layer.
• The report core contains a summary, a per-property snapshot, migration recovery trends and an action plan. Other sections can appear when the data warrants them.
• An SEO specialist reads and approves every report before it reaches the client.
• No false or ambiguous signal has been reported so far. The observed problem was different: some versions supplied too much information, explained it too densely and needed simplification.
• Automation should separate an observation, a possible explanation, a required verification and a proposed action.
• Native Search Console limits, including anonymised queries, row truncation, aggregation and potentially incomplete recent data, must remain visible.
• The specialist remains accountable for validation, priority and any consequential change to the website.

The system prevents a manual assembly role from forming. It does not remove expert reading or care.

In this article

1. Automate repetition, not uncertainty
2. The four layers of the workflow
3. The contract for an actionable report
4. Why Search Console is never the complete view
5. Track change without manufacturing alarms
6. What AI can recommend, and what it cannot conclude
7. When to automate the reporting, and when to keep it simple

1. Automate repetition, not uncertainty

A migration changes the very elements that shape how a website is read in organic search: URLs, hierarchy, templates, internal links and sometimes content. Monitoring therefore cannot stop at a snapshot. It needs to show a recovery trajectory. What to protect before and during the migration is covered in website redesign or migration: how to protect your Google visibility; this post starts where that one ends.

In this case, the team had no historical manual report. The decision was to automate from the outset the two families of work that would otherwise recur.

The first is repeatable:

• retrieve the same categories of data each week;
• apply the rules defined in the reporting brief;
• compare the current snapshot with the complete previous history;
• calculate changes;
• produce the same chart types;
• arrange the output in a stable report structure.

The second requires judgement:

• decide whether a change matters to the business;
• distinguish seasonality from an incident;
• relate a shift to a release, migration or external event;
• choose between a technical change, an editorial change or no action;
• decide how much time and budget to commit.

The workflow automates the first family and prepares the second. An SEO specialist approves the result before it leaves the team. That separation was designed into the system rather than added after an incident.

That boundary creates a better report. Instead of hiding uncertainty beneath a definitive recommendation, the system can say: this changed, these explanations remain compatible with the data, and this verification would separate them.

2. The four layers of the workflow

The public architecture has four layers.

1. Collect

The system retrieves the metrics and dimensions required for the defined property and period.

Google’s Search Analytics API accepts date ranges, dimensions, filters, search type, aggregation and pagination (official documentation). A workflow can therefore repeat a structured extraction without asking somebody to rebuild the query in the interface.

The exact metrics, dimensions and windows used in this project are not public. We do not claim that one reporting template suits every website.

2. Compare

Every weekly snapshot is placed within the full available history: one property over time, one page or group, a changing distribution, or movement inside a defined segment. The report therefore asks not only “what changed since last week?” but “where does this week sit in the recovery since migration?”

The comparison needs to preserve context. Comparing a promotion week with an ordinary week can produce an accurate difference and a useless conclusion.

3. Represent

The workflow creates the charts required by the report. A useful visual retains the unit, period, segment and reference window. Without them, a chart becomes persuasive decoration.

4. Interpret and prepare the next step

The reporting brief sets the rules and expected structure. An LLM layer then summarises the changes, assembles the relevant sections and proposes recommendations. The output should keep what was observed separate from what still needs to be checked.

Search Console
      │
      v
parameterised extraction
      │
      v
comparison with the full history
      │
      v
charts + changes
      │
      v
observation ─> hypotheses ─> verification ─> proposed action
      │
      v
automatic weekly report
      │
      v
SEO specialist approval

3. The contract for an actionable report

An automatic report does not become useful simply because it contains more prose.

The public core of this post-migration report has four blocks.

The structure is not rigid. The workflow can add a section when that week’s data warrants it. This is the opposite of padding a report so it looks complete: a block exists because it helps somebody understand or decide.

For every important signal, we recommend four distinct fields.

This structure prevents “traffic is down, rewrite the page” from travelling through the report without an evidence step in between.

It also makes the output usable across roles. The SEO lead sees the signal. A developer sees the technical check. The content owner sees the editorial question. The decision-maker can see what is known and what remains an explanation.

4. Why Search Console is never the complete view

Automating a source does not expand its coverage.

Google documents several important limits:

• some queries are anonymised to protect privacy;
• report tables do not expose every row and may be truncated;
• data is aggregated according to the property, canonical page and requested dimensions;
• recent data may be incomplete and continue to change;
• Search, News and Discover use separate reports or types.

The API returns at most 25,000 rows per request before pagination and can identify the point at which recent data becomes incomplete (Search Console API). Google also states that bulk export to BigQuery contains the performance data available for the property except anonymised queries (Search Console Help).

These are not technical footnotes. They govern what the report is allowed to say.

An automated extraction should retain:

• source and property;
• period and relevant time zone;
• filters and dimensions;
• type and aggregation;
• final or incomplete data state;
• known limits that affect interpretation.

Without that context, a recurring report simply repeats the same ambiguity on schedule.

5. Track change without manufacturing alarms

The workflow runs once a week. On every run, the current snapshot joins the history and is compared with all earlier snapshots. That prevents post-migration recovery from being reduced to an arbitrary contest between two weeks.

But “it moved” does not automatically mean “act now”.

A change can be:

• real and important;
• real but expected;
• exaggerated by a small base;
• caused by the chosen comparison window;
• based on recent data that is still incomplete;
• isolated to a segment that does not affect the wider priority.

The system applies the parameters defined in the brief, exposes the comparison base and avoids turning every difference into an emergency. The exact metrics, dimensions and thresholds remain confidential.

No false or ambiguous signal has been reported in the use described by the team. The observed weakness was editorial: some versions contained more than the reader needed, explained it too densely and created confusion.

That feedback changed the quality question. Finding variations was not enough; the report also had to prioritise, compress and write for the person who would read it. In automated reporting, overload is a failure mode too.

The weekly cadence belongs to this case. It is not a universal recommendation: a daily report is not more mature than a monthly one if nobody can act at that frequency.

6. What AI can recommend, and what it cannot conclude

In this workflow, the reporting brief sets the rules. The LLM layer then carries out the analysis and drafting automatically. It can:

• summarise the most visible changes;
• group signals according to defined parameters;
• propose explanations compatible with the data;
• prepare an initial verification order;
• draft the report around its core structure;
• add a section when that week’s data warrants it.

Search Console alone cannot establish:

• the certain cause of a decline;
• the quality of the post-click experience;
• the number of qualified leads or resulting revenue;
• the isolated effect of one page change while other factors moved;
• that a generic recommendation is the business priority.

Even Google’s native Search Console recommendations are optional, change over time and appear only when Google identifies something it considers actionable (Google Search Console Help). They are suggestions to examine, not a universal causal diagnosis.

AI should reduce the cost of preparing judgement. It should not give the report confidence that the source does not possess. For the manual version of that diagnosis, see why did your website traffic drop?.

The human control is concrete. Before the client receives the document, an SEO specialist reads it, checks that the conclusions remain proportionate to the data and simplifies passages that are unnecessarily complex. In most reported cases, this takes 15 to 25 minutes.

That figure describes the current review time. It is not a before-and-after saving because no earlier manual reporting process existed.

7. When to automate the reporting, and when to keep it simple

This workflow is useful when:

• a migration or rebuild has changed the website structure;
• the complete trajectory matters more than a simple comparison with the previous period;
• the same extracts and comparisons recur;
• multiple properties, segments or periods make manual review costly;
• the team spends time rebuilding charts and report structure;
• parameters and data limitations can be documented;
• a specialist can approve the report before distribution.

It is probably excessive when:

• one person can answer the need in the native report in a few minutes;
• no recurring decision depends on the output;
• volumes are too low for fine-grained changes to be meaningful;
• the team wants automatic prose but has no action attached to it;
• nobody will investigate important anomalies.

The question is not “can the report be automated?” Most of it can. The useful question is:

Which recurring decision should this report make faster or safer?

Conclusion: 100% automated, not 100% certain

Every week, this system automatically collects, compares the latest snapshot with the full history, visualises, analyses and produces the report.

That automation is real. It does not give Search Console information it does not contain, and it does not turn correlation into cause.

The right output is not a confident caption beneath every chart. It is a readable chain: observation, possible explanation, required verification and proportionate action.

In this case, automation did not replace an old manual report. It established post-migration monitoring without creating that routine. The specialist spends their time reading, simplifying and approving the result, where their judgement has real value.

Frequently asked questions

Sources and limits

• The client workflow is described from IZZY’s own project knowledge, extended on 7 September 2026. Havana Club is named; its properties, exact parameters, thresholds, reports and logs remain confidential.
• Search Analytics: query, official Search Console API documentation.
• Performance report dimensions and data groupings, bulk export to BigQuery and Search Console recommendations, Google documentation.
• Google’s documentation supports the mechanics and limits of the source; it does not prove the effectiveness of the client workflow.
• The 15–25-minute review time is a reported observation for most reports, not an independently measured saving. No detection rate, traffic improvement, conversion or revenue result is claimed.