📘 Use this page when a Botify MCP answer is empty, incomplete, or does not look right. Most of these are configuration or phrasing issues rather than faults, and each one has a check you can run yourself.
Overview
When an answer comes back empty or thin, the cause is usually a missing prerequisite on the project rather than a problem with the connection. When an answer comes back full but wrong, the cause is usually the question: an unstated level of detail, an undefined term, or an assumption about which data was used. This page covers both.
Prerequisites
Confirm that Botify MCP is connected to your AI assistant and that your account has access to the project you are asking about. If either is missing, no question returns useful data.
When a question returns little or nothing
Different questions depend on different data being set up on the project. Check the row that matches what you asked for.
What you need | How to confirm it | Questions that depend on it |
A completed crawl | Your project shows a recent SiteCrawler snapshot | Crawl and indexability questions, structured data checks, migration mapping, cannibalization, facet work, extraction checks |
Log files being ingested | Monthly log partitions are present and current | Bot behavior questions, and the crawler-access half of any AI visibility question |
A connected Search Console property | Your project configuration lists the property | Trend and ranking-decay review, cannibalization, metadata testing baselines |
AI Visibility configured | Prompt runs are returning results | The citation half of any AI visibility question |
PageWorkers enabled | Your project configuration reports PageWorkers status | Metadata testing, facet work |
Annotations kept up to date | Your project returns more than a handful of annotation entries | Timeline correlation, which has nothing to correlate against otherwise |
Revenue per visit stored on the project | Your project metadata holds a revenue-per-visit figure | Opportunity sizing, and any question that converts traffic to value |
Your expected field values, written down | Defined before the run, not after | Extraction checks, structured data validation |
A second connector with write access | Your repository, issue tracker, or CMS is reachable from the same AI client | Anything that files a ticket or opens a pull request |
When the answer is full, but looks wrong
The numbers are aggregated at the wrong level
Search Console performance data is recorded at the URL, keyword, device, country, and day level. If you do not name the level you want, the answer may total across a dimension you meant to keep separate. Ask again, naming the level explicitly.
The numbers do not match your analytics platform
Your analytics platform and your server logs both measure something they call visits, and they measure different things. Neither is wrong. Ask which source a figure came from, and decide which source is authoritative for that metric before the number reaches a report.
The same question returns a different answer on a second run
Answers read from live search results, AI Mode, or answer engines query those surfaces in real time, so the same question asked twice can return different results. Answers read from crawl, log, or Search Console data come from a stored snapshot and reproduce until the next crawl or log ingestion. Check which kind you asked for. Record the date and time alongside any live reading you intend to keep.
A total covering several projects looks wrong
An answer spanning several projects is built by listing every project your account can reach and querying each in turn. That list can include test crawls, ad-hoc experiments, and abandoned projects alongside your real properties, and projects are configured unevenly: log ingestion, Search Console, AI Visibility, and stored revenue per visit vary from one to the next. A combined total can therefore be a partial one. Ask which projects the answer covered, and whether each one had the data behind it.
A correlation looks confident but explains nothing
Correlating a traffic or ranking shift against site changes depends on your annotation log being current. Against an empty or stale log, the answer still looks confident and has no explanatory value. Check the log before relying on the result.
The answer covers only part of what you asked
Given a broad structure such as a whole navigation tree, the assistant may work one branch thoroughly and present that as the complete answer. Name the sections you want covered, or check the coverage before trusting the list.
Extraction results look complete but are not
Without a defined set of expected field values, there is nothing to test an extraction against, and a silently partial extraction reads the same as a complete one. Define the expected values first, then ask for the failures rather than the full result.
The output is hard to read
Ask for the output format you want when you ask the question. Left open, a large analysis can come back as a spreadsheet of dozens of separate tables that is harder to read than the underlying data. Once you have a result, interrogate it rather than re-running the whole analysis: asking about one specific item, or whether two items are the same problem, is faster than starting again.
The answer is confident but wrong
An AI assistant states incorrect answers with the same confidence as correct ones, and it cannot check itself against a second source the way you can inside the platform. Verify anything you plan to act on or pass to someone else, and ask where a number came from. Asking what the assistant could not verify is as useful as reading what it flagged, because an unverified claim reads the same as a confirmed one.
Read next
Related topics