Marketing Analytics Gets a Chat Box: Which AI Answers Can an Analyst Check?

All 3 of the analytics platforms most marketing teams use now answer questions typed in plain English. Google calls its version Ask Advisor, Amplitude calls it Global Agent, and Mixpanel renamed Spark to Mixpanel Agent in June 2026. For marketing analytics, the price is not the difference. The difference is whether your analyst can open the query that produced the number before it reaches a board slide.

The short answer. GA4, Amplitude and Mixpanel all offer natural-language questions on their entry plans. Mixpanel documents the clearest audit path, because its assistant writes a standard report you can open and edit. Amplitude links every claim to a chart. GA4 links to reports but documents no way to see the query behind an answer.

What the chat box in marketing analytics actually does now

The first generation of these features translated one sentence into one chart. The current generation describes itself as an agent. That word matters here because it means the assistant plans several steps, searches your saved work, and in some cases edits it.

The useful way to compare them is not by how clever the answer sounds. It is by what object the answer is attached to. A sentence in a chat window is a claim. A saved report with a visible event, filter and date range is evidence.

Google: Ask Advisor inside GA4

Google announced Analytics Advisor in November 2025 alongside Ads Advisor, and said both would reach English-language accounts globally in early December. The announcement said it would cover both Standard and 360 properties. The current help page names the feature Ask Advisor and still labels it beta.

Google's examples are a mix of data and setup questions. Trends in active users, why revenue dropped on a given day, how to create an audience, and what your measurement ID is. Google says the assistant performs a key driver analysis to explain spikes and drops.

Amplitude: Global Agent, labelled Ask AI

Amplitude's Global Agent documentation describes it as having access to your entire project. It discovers events and properties, finds existing dashboards and analyses, and explains your data taxonomy. It can also build charts, create cohorts, and edit dashboards and notebooks on request.

Amplitude warns that the agent may take 30 to 60 seconds to research an answer, and that answers default to 2 to 5 sentences. That is a sensible default. Short answers backed by a chart are easier to check than long ones.

Mixpanel: Mixpanel Agent, formerly Spark

Mixpanel launched Spark in July 2023 and renamed it Mixpanel Agent in June 2026. The current documentation says it creates new Insights, Funnels, Retention and Flows reports from a description. It also adjusts metrics, filters, breakdowns and time ranges on existing reports.

The agent FAQ says Mixpanel Agent is powered by Claude. The 2023 launch coverage described a different model, which is a useful reminder that the engine under any of these features can change without the feature name changing.

Which plans include AI analytics tools, and what they cost

This is the part of the comparison that has collapsed. In each case the natural-language feature is bundled into the plan you already pay for, including the free tiers. None of the 3 vendors sells it as a separate AI add-on in the pricing pages and documentation reviewed here.

Every cell is taken from the vendor's own help pages, pricing pages or announcements as read on 20 October 2026. "None found" means the pages reviewed did not state a cap, not that no cap exists.

The entry price, where vendors publish one

Amplitude's free plan covers 2 million events a month. Its Plus plan starts at $0 with the first 2 million events free, and Growth and Enterprise are custom quotes on event-based pricing.

Mixpanel's free tier covers 1 million events a month. Its Growth plan is the only one of the 3 with a public rate card, shown below. GA4 does not appear in the chart because Google's announcement and help pages publish no price for the feature itself.

Why bundling changes the buying question

When a feature costs nothing extra, the procurement conversation skips it. That is the risk. Your team gets an assistant that can write and edit reports without anyone deciding who checks its output.

I would treat the absence of a price as a reason to write the review rule now, before usage spreads. The cost of an unchecked number is not on any of these rate cards.

What each AI analytics tool can query

Scope decides which questions are safe to ask. An assistant that sees one property cannot explain a drop caused by a change in another system, however fluent the answer sounds.

GA4 sees one property, including its settings

Google states that answers about Analytics data rely only on information from within that specific property. That is a clean boundary. It also means the assistant cannot see your CRM, your ad spend outside Google, or the offline pipeline that closed the deal.

The setup questions are where Ask Advisor looks strongest. Finding a measurement ID or the steps to build an audience is low risk, easy to verify, and genuinely faster than searching the help centre. For the gaps a single property cannot fill, see the analysis of attribution blind spots in dark discovery.

Amplitude and Mixpanel see the governed taxonomy

Both product analytics vendors give the agent the project's event and property definitions, and both say it only surfaces data the requesting user can already see. Amplitude says agents honour existing access controls on projects, dashboards and datasets. Mixpanel makes the same commitment on its AI compliance page.

Mixpanel adds a quality signal. Results that draw on events not yet reviewed in Lexicon, its data dictionary, are flagged as unverified. That is the right instinct: it tells the reader the answer rests on data nobody has signed off.

The query behind the answer: the test that separates them

An analyst checking an AI answer needs 4 things. The event or metric used, the filters, the date range, and the comparison baseline. If any one is hidden, the analyst is checking the prose, not the number.

Mixpanel writes a report, so the work is visible

Mixpanel's position is stated plainly in its Spark launch post. A report built by the assistant is viewable and editable like any other, so you can open the query builder and see which events it used. The same post commits that any generative feature it ships will be able to "show its work".

This is the design I would choose. The assistant produces the same artefact a human analyst would, so the existing review habit still applies. Nobody has to learn a new way to audit.

Amplitude links every claim to a chart

Amplitude's documentation says every claim includes a link to supporting data, and that chart previews can be opened for full chart details. Since an Amplitude chart is itself a saved definition, opening it should show the event and segment settings. The documentation describes the link, not the inspection step, so test it on your own project.

The weaker point is write access. Amplitude says edits to dashboards and notebooks apply immediately to saved content. A colleague's dashboard can change because someone asked a question in plain English, and the review happens after the fact, if at all.

GA4 links to reports and explains drivers

Google promises visualisations and links to helpful reports. A link to a standard report is useful when the answer is a simple trend. It is much less useful for the key driver analysis Google highlights, because a list of reasons for a revenue drop is a statistical judgement, not a report row.

My view is that GA4's assistant is fine for orientation and setup, and should not be the source of any number you repeat to a client without rebuilding it in a report or exploration yourself. That is a judgement on the documentation, which says nothing about how the driver analysis is calculated.

How each vendor documents accuracy and limits

Read the three caveats side by side and the difference in tone is striking. Two vendors tell you the assistant can be wrong. One tells you it does not guess.

"Never fabricates" is a promise nobody can keep

Amplitude's documentation says the agent never fabricates or guesses and reports what the data shows. The same page tells users to always review and validate AI-created analyses before important decisions. Both cannot carry equal weight.

I trust Mixpanel's wording more, precisely because it concedes the failure mode. A vendor that says its model can be inaccurate has told you to keep checking. A vendor that says it never guesses has given a hurried marketer permission to stop. The pattern is familiar from the hallucination audit of deep research tools, where confident sourcing and correct sourcing were different things.

No vendor has published a test set

None of the three documentation sets reviewed here reports a measured accuracy figure, a benchmark of questions with known answers, or an error rate by question type. That is the number a buyer needs most, and it is absent across the category. Your own team can build a small version, as described in the guide to building an LLM evaluation suite.

Data handling and admin controls compared

Each assistant sends some of your analytics data to a large language model. The vendors differ on what they disclose about that path, and on who can switch it off.

Amplitude publishes the most granular control. Its AI Controls page lists three states: off, admin preview, and on for the organisation. Global Agent is on by default, and individual users cannot override the organisation's choice. Amplitude also says that for customers provisioned in its EU data centre, data never leaves the EU when AI features are used.

Mixpanel's AI compliance page says AI features are on by default, except for Enterprise customers in its EU or India data residency programmes and those who have signed a Business Associate Addendum. The reason is stated on the same page: for residency customers, inputs and outputs may be sent to providers outside the EU or India. An admin toggle switches AI off for everyone in the organisation.

Google's data page is clear on training and retention, but the pages reviewed here describe no admin switch for Ask Advisor. If your contracts restrict where analytics data is processed, that gap is worth raising with your Google account team. For a wider comparison of retention terms, see the ranking of AI vendor data retention policies.

The MCP door is separate from the chat box

Both product analytics vendors now expose their data to outside assistants through the Model Context Protocol. Mixpanel says MCP is on by default for Free and Growth and off by default for Enterprise, where an admin must enable it. Amplitude lists MCP on every plan. That means your data can be queried from tools your analytics admin does not manage, a trade-off covered in the explainer on MCP as an interoperability standard.

Where this comparison is weakest

This post compares documentation, not behaviour. That is a real limit, and it cuts in a specific direction.

No live property was tested for this analysis. A vendor that documents its audit path badly may still have a good one in the product, and a vendor that documents it well may ship something rougher. Amplitude's chart links in particular may expose full definitions in practice. The documentation simply does not say so in those words.

The GA4 picture is the thinnest. Standard and 360 parity rests on Google's November 2025 announcement rather than the current help page, and the help page does not define which accounts are eligible. Google could also add query visibility at any point while the feature remains in beta.

There is also a fair argument against the whole test. Most marketers asking "which pages lost traffic last week" do not need an auditable query. They need a fast pointer to the right report, and GA4 does that. Demanding analyst-grade provenance for orientation questions adds friction for little gain.

I accept that for orientation. I do not accept it for any number that leaves the analytics tool. The test in this post is meant for the second case, which is where a wrong figure becomes expensive.

Frequently asked questions

How to set a review rule this week

Take one question your team answered by hand last month, with a number you already trust. Ask it in whichever assistant you have, then open whatever the answer links to and compare event, filter, date range and baseline. Write down which of the four you could see. That single comparison tells you more than this post can.

Then write one sentence into your reporting process: any figure produced by an assistant is rebuilt as a saved report before it goes to a client or a board. If you use Amplitude, add a second line covering who may let the agent edit shared dashboards. The interface is shifting toward conversation, as argued in the post on software after the dashboard, and the saved report is what keeps that shift checkable.

Related reading