You can ask Google Analytics questions with AI, but you still need to check its answers before making decisions. Define the period, metric and comparison, then compare the result with reports and business records. Google's conversational API is in alpha. It cannot repair missing events, measurement errors or causal conclusions that the data does not support.
What has actually changed in Google Analytics?
Many businesses have traditionally explored their data through standard reports, explorations or dashboards prepared by someone familiar with analytics. If a marketing manager wanted to understand why a campaign attracted visits but few enquiries, someone needed to identify the appropriate dimensions, metrics, filters and periods. The conversation often began outside the tool: could someone produce this report? The work involved defining what the number meant, rather than simply retrieving it.
Google's documentation records the addition of properties.chat to the Data API v1alpha on 14 September 2026. The method accepts a natural-language question, can continue a conversation using a session identifier, and returns text or table blocks. It relates to Ask Advisor, which is also documented as an experience within Google Analytics. The existence of an API does not mean every property, integration or external supplier offers the same experience. Check availability and access in your actual environment.
For a small business, the opportunity is to make data easier to discuss for whoever decides promotions, reviews content or coordinates sales. The limitation matters just as much. A conversational tool works with what has been measured, including the gaps and mistakes. If a form fails to send its event, or consent settings prevent particular collection, a well-written question cannot recover the missing information.
Which business questions should you start with?
Start with a decision instead of asking the system to analyse the whole business. A shop might examine product landing pages reached through organic search and compare their recorded actions. A services business might review visits to its contact page by channel. A school might inspect the journey towards an enquiry form. These are possible questions, not measured results from meiika clients.
A useful question specifies four things: the period, segment, metric and comparison. For example: compare organic-search sessions on mobile and desktop during the last complete month and show the landing pages. Then ask for a table, an interpretation and uncertainties separately. If the question mixes users, sessions, conversions and revenue without definitions, an answer may sound convincing while remaining difficult to check.
Negative questions are helpful too. Ask what cannot be concluded from the available data. That draws attention to missing information about calls, returns, campaigns or enquiry quality. Fewer sessions alone do not demonstrate lower sales. More completed forms do not guarantee more customers. Analytics describes part of the behaviour that was recorded; it does not automatically explain every cause behind it.
Small teams can keep a shared set of questions about acquisition, page performance, forms and changes against the previous period. Avoid turning every query into another KPI. Maintain a few stable definitions and document why the comparison matters. That continuity makes the next meeting useful, rather than producing a growing collection of screenshots without context or an agreed action.
How do you check an AI answer before using it?
First, check meaning. Has the tool understood users or sessions? Which property time zone is involved? What events does it treat as relevant? Are the periods the same length? Next, check the technical result against a conventional GA4 report, or an equivalent reporting API query where possible. Finally, compare the observation with orders, enquiries or CRM records. Different systems do not necessarily match event for event, but unexplained differences need investigation.
Google warns in the properties.chat reference that its AI output can contain inaccuracies. Keep the exact question, period, dimensions and consultation date beside the conclusion. If the result informs a budget decision, record what was verified and what remains a hypothesis. Do not automatically attribute a change to a campaign when seasonality, tagging or website changes might also explain it.
When a number looks strange, check the foundations: the installed tag, consent, duplicate events, UTM parameters, key events and exclusions. If a landing page has changed address, confirm that visits are grouped as expected. When comparing mobile and desktop, verify that the metric is comparable. Natural-language questions become more useful when measurement is reliable; they do not replace that foundation.
Avoid sharing screenshots containing sensitive information unnecessarily. If a summary goes to a team channel, decide who should receive it, how long it stays available and which details it can expose. A report suitable for management may be unsuitable for a discussion open to everyone. Easier access to information still needs an access policy and a clear purpose.
When is a conversational report worth automating?
Automation can help when the same question comes up every week and someone will review the result before acting. An internal report could prepare a channel comparison and a table for the marketing manager to inspect for anomalies. The official reference describes text and table blocks, a sessionId for related questions, and hourly and daily token quotas that can be requested in the response.
Separate extraction, interpretation and decision. Retrieve the data first, prepare a commentary identifying possible changes second, and leave the decision about investigation to a person. Connecting an AI conclusion directly to increased advertising expenditure or switching campaigns off is a poor starting point. The tool can help prioritise questions; financial actions need additional evidence and an accountable decision.
Define what happens if the API returns an error, an empty response or a table missing expected columns. Keep an unverified state rather than silently substituting zero. Zero is a value; missing data is something different. Record the query and the version of the logic used to interpret it. This makes it possible to reconstruct where a recommendation came from instead of relying on someone's memory.
You do not need to begin with an autonomous agent. A simple flow preparing a summary for review may be more useful and easier to supervise. Technical complexity only earns its place when it solves a repeated need. Before building, name the reader, the next action and the evidence that would make the answer trustworthy. Automating a report nobody uses merely adds another item to the inbox.
Which permissions and limits need checking?
The Google reference requires the OAuth scope https://www.googleapis.com/auth/analytics.chatbot.read for properties.chat. This differs from permissions for other reporting operations. The person or application also needs appropriate property access. Do not copy credentials into shared documents or grant administrative privileges simply because that is convenient. The technical owner should manage tokens, secrets, access and revocation.
The method is documented in v1alpha. Before making it part of a critical process, check availability, stability, quotas, errors and documentation changes. A demonstration with one successful question does not establish that the integration handles every case. Include periods with no data, custom event names, ambiguous questions and failed authorisation in the test plan.
Interpretation has limits too. A question about why enquiries fell may produce a plausible explanation without proving causality. Ask for observed facts, hypotheses and checks still needed to be distinguished. Keep references to the data consulted. Do not publish commercial figures or claims about customers unless they have verifiable support.
| Check | Record | Stop and review when |
|---|---|---|
| Question | Period, segment and metric | Definitions are ambiguous |
| Answer | Table and interpretation separately | The underlying data is missing |
| Comparison | Report and business record | A difference remains unexplained |
| Action | Owner and reviewed evidence | Only a hypothesis is available |
How have we approached it at meiika?
Our approach to measurement and Growth & MarTech starts with questions that support a decision. Before automating a summary, identify the available data, who will review it and which conclusions cannot yet be supported. A useful discussion makes uncertainty visible instead of hiding it behind fluent prose.
We also study custom systems that connect information with human review. These are design criteria, not a dashboard implemented for a named client or a claim of measured customer results. The right first step depends on the questions your team actually needs to answer.