All guides

Ask ChatGPT or Claude about your website traffic: questions that work and questions that mislead

Last checked:

Start with three questions that name the site, the period and the number you want. They give answers you can check against the dashboard. Questions about causes, individual people, AI citations or totals built from a top list produce confident answers without data behind them.

“List my sites, then show visitors, pageviews and traffic sources for [site] for the last 30 days, with the exact dates the period covers.”
“Show the ten most visited pages on [site] for the last 30 days together with the site total for the same period, so I can see how much of the traffic the list covers.”
“Which goals and funnels are set up for [site], what does each one count, and how many completions did each get in the last 30 days?”

Connecting takes one step: add Savri from the directory in your assistant and sign in with your analytics account. The connection guides cover ChatGPT, Claude and the manual route. This guide is about what to ask afterwards.

How we tested the questions

On 12 October 2026 we ran eighteen candidate questions through the published connection against one of our own sites, and one more of our sites for the Bing reports. For each question we noted which tools the assistant would call, whether the returned data supported the answer, and what went wrong when it did not. The connection reads the same reports you see in the dashboard: visitor statistics for rolling 7, 30 or 90 day periods, pages, sources, countries, goals, funnels, event properties, and Google and Bing search reports where those are connected. It does not read raw visitor records.

The assistant decides which tool to call and how to phrase the result. The tools return structured data with their own warnings, for example that absent rows are unknown rather than zero. A good question makes the assistant use the right tool and repeat those limits. A bad question invites it to fill the gap.

Questions that work

1. Visitors and pageviews for a period, with a comparison

“How many visitors and pageviews did [site] have in the last 7 days compared with the previous 7 days? State both date ranges.”

What the assistant reads: the overview for a rolling period and the previous period of the same length. Visitors are daily identifiers based on a hash, not people, so the same person counts again the next day.

What you can use it for: spotting a change worth looking into. In our test the overview came back with both periods and the change in percent.

How to check it: open the site overview in Savri with the same period. The figures should match. A percentage from a small base swings a lot; ask for the absolute numbers when the counts are small.

2. The most visited pages, next to the site total

“Show the ten pages with the most pageviews on [site] in the last 30 days and the total pageviews for the site in the same period.”

What the assistant reads: a top list of pages and, from the overview, the site total.

What you can use it for: seeing where attention goes. In our test the ten pages covered most, not all, of the pageviews; the rest was spread across pages outside the list.

How to check it: the Popular pages card in the overview. If the assistant presents the sum of the list as the site total, that is the trap in the next section.

3. Where visitors came from

“List the traffic sources for [site] in the last 30 days with visitor counts, including every source label you receive, not only the top five.”

What the assistant reads: the sources report. A source is evidence that a visit arrived with that referrer or campaign label.

What you can use it for: seeing whether search, direct traffic, a campaign or an AI service brought the visits. In our test visits from chatgpt.com and claude.ai appeared as ordinary sources. Search traffic appeared under two labels, the referrer and a campaign source, which only a complete list shows.

How to check it: the Sources card in the overview. A visit from an AI service is a referred visit, not proof that the service recommended you.

4. Goals and funnels, with their definitions

“Which goals and funnels exist for [site]? For each one, tell me exactly what it counts and how many completions it had in the last 30 days.”

What the assistant reads: the goal and funnel definitions and their completion counts. A goal counts a page visit or a named event. A funnel counts sessions that pass its steps in order.

What you can use it for: knowing what the numbers mean before you compare them. An outbound click is a click, not a sale.

How to check it: the Goals page in Savri, where each goal shows its definition and “Test match” lists matching events from the last 30 days.

5. Google clicks and impressions per page

“From Google Search Console, which pages on [site] received the most clicks in the last 30 days of final data? Include impressions, and tell me the last date the data covers.”

What the assistant reads: the Google page report through the site’s Search Console connection, which is available from the Basic plan. Google data is final with a delay: in our test the latest final date was two days before the day of the question.

What you can use it for: finding pages that earn search clicks. In our test the clicks per page added up to the overview total.

How to check it: the Google view in the site dashboard, or Search Console itself, with the same dates. Lower position is better, and positions should not be averaged without weighting.

6. Bing search queries from the latest weekly report

“From Bing, show the latest weekly query report for [site]: the report date, the queries with clicks and impressions, and whether the report is complete.”

What the assistant reads: Bing’s weekly Web report for your own Bing connection, labelled with the report date Bing uses. Daily Bing traffic is a separate report that covers all Bing surfaces.

What you can use it for: seeing which queries show your pages in Bing. In our test the tool picked the most recent report label itself and said the report was not complete.

How to check it: the Bing view in the site dashboard. Weekly report rows should not be added to the daily totals; the tool says so in its response.

Questions that mislead, and what to ask instead

7. “How many pageviews did my site get? Add up the top pages.”

What happens: the top list is a sample of the busiest pages. Its sum is lower than the site total, and an assistant that adds the rows reports a number that is wrong but plausible. The same applies to Google queries: Google returns top rows and omits anonymised queries, so the rows never add up to the total.

“What was the total number of pageviews on [site] in the last 30 days, taken from the overview, not from a list of pages?”

8. “Why did traffic fall on [date]?”

What happens: the tools return counts per day, source, page and country. None of them contains a cause. The assistant can still write a fluent explanation, such as a campaign ending or a search update, with no data behind it.

“Compare [site] for the 7 days ending [date] with the 7 days before. Show visitors per day, the sources and pages that changed most, and list what the data cannot explain.”

9. “Which search queries bring the most Google clicks?”

What happens: in our test the overview reported a small number of clicks in the period, while every query row the tool returned showed zero clicks. The clicks came from queries Google anonymises and never returns. An assistant that reads only the rows would say that no query brings clicks.

“From Google Search Console for [site], show the total clicks for the last 30 days of final data, then the returned queries sorted by impressions, and state that absent queries are unknown rather than zero.”

10. “Why does the funnel show zero sign-ups when the sign-up page has visitors?”

What happens: a funnel counts sessions that pass its steps in order. In our test the sign-up page had visitors in the period, but the funnel’s third step showed zero, because those sessions had not visited the two earlier steps first. The last step was an event that may no longer be sent at all. The question presumes a fault; the data shows a definition.

“Show the exact steps of the funnel on [site], the sessions at each step for the last 30 days, and separately how many visitors reached the sign-up page in the same period. Do not explain the difference, just show both.”

11. “How many Bing clicks do I have?”

What happens: without dates, the daily Bing report in our test returned the whole available history, more than a year of days, and the total covered all of it. A 30 day question gave a very different number.

“From Bing daily traffic for [site], show clicks and impressions from [first date] to [last date], and tell me which dates in that range have no data.”

12. “Is my site cited by ChatGPT?”

What happens: no tool measures citations. The sources report shows visits that arrived from chatgpt.com, which is a different thing. The Google and Bing tools say in their responses that they carry no AI citation metrics. An assistant may still answer with a yes based on the referral count.

“How many visitors arrived at [site] from chatgpt.com and claude.ai in the last 30 days? Treat them as referred visits only, and tell me that citations are not measured here.”

13. “Who visited from Germany last week?” and “Did a specific customer visit?”

What happens: there is no tool for individual visitors. Visitor identifiers are daily hashes and are not exposed. The honest answer is a country count. Anything about a person is invented.

“How many visitors did [site] have from each country in the last 7 days? Do not attempt to identify anyone.”

14. “Did Google clicks drop yesterday?”

What happens: Google data arrives with a delay and the tool reports the latest final date. In our test that date was two days back, so “yesterday” did not exist in the data. An assistant that ignores the date reads the last available day as yesterday.

“From Google Search Console for [site], what is the latest final date, and how did clicks per day develop over the last 7 final days?”

Numbers to be careful with

  • Percentages from a small base. A 90 day comparison on a site that started measuring recently showed a change of several hundred percent in our test. Ask for absolute numbers.
  • Very small counts. A breakdown of a custom event property returned values with two events each. Ranking them says nothing. Ask whether the counts are large enough to rank.
  • Periods the tools do not offer. Visitor statistics come in rolling 7, 30 and 90 day windows, not calendar months. Google supports exact dates within roughly 16 months. Bing weekly reports use Bing’s own report labels. If you ask for “last month”, ask the assistant to state the dates it actually used.
  • Not connected is not zero. For a site without a Bing connection, the tool returned “not connected” and a setup link. That is a missing report, not zero clicks.
  • Configuration is not measurement. A goal with zero completions can mean nobody acted, or that the event is not being sent. In Claude, the connector can run a measurement check for the site; in ChatGPT, use the measurement check in the site dashboard. Goals count from the moment they are created.

What the assistant cannot know about your site

It does not know what you changed and when, what a campaign cost, whether a form message was delivered, who a visitor is, whether an AI answer mentioned you, or why a number moved. It knows what the connected reports returned for the period it asked about. Keep the dates and the source of every figure in the answer, and open the dashboard when a conclusion matters.

This applies to any analytics tool with an AI connection, not only Savri. The report names, periods and limits above are ours; the habit of asking for dates, totals and what is missing works everywhere.

Sources

OpenAI, Connected apps in ChatGPT, read 12 October 2026: apps are added under Settings and Plugins, availability depends on plan, region, workspace, role, model and interface. Anthropic, Use connectors to extend Claude’s capabilities, read 12 October 2026: connectors are added from the Connectors Directory, Claude inherits your permissions from the connected service, and free accounts are limited to one custom connector. On the same day the ChatGPT directory listed Savri version 1.1.0 and the Claude directory connector exposed the server’s 30 tools. The test figures come from our own sites and are not published. Tool limits are quoted from the tools’ own responses.