Google Search Console and AI: AI Overviews, AI Mode, the Generative AI Report and MCP

What Search Console tells you about AI search now

Search Console now covers Google's AI search in three ways. AI Overviews and AI Mode clicks sit inside your normal Web data, a separate generative AI report counts impressions, and a setting lets owners exclude a site from AI features altogether. The timeline is short. On June 3, 2026, Google introduced generative AI performance reports for Search and Discover, first for a subset of sites, and on August 31, 2026 they reached all websites worldwide. The Search generative AI control arrived with the same rollout. What Search Console lacks is an official connection to Claude or ChatGPT. Google publishes no Search Console MCP server, so community and vendor servers fill the gap by wrapping the same API.

How AI Overviews and AI Mode show up in your data

One position, one click

An AI Overview occupies a single position in the search results, and every link inside it gets that same position. Clicking an external link in an AI Overview counts as a click ^[1]. AI Mode follows the same logic: a click on an external page counts as a click, and position follows the same methodology as an ordinary results page. So AI traffic already sits inside your Web clicks today. There is no separate click total. Experiments in Search Labs are not in the data at all, because Google still treats them as in active development. If you saw your site in a Labs feature, Search Console will not show it.

Why AI Mode creates strange queries

In AI Mode, a follow-up question counts as a new query. A single session where someone asks three follow-ups can therefore produce three separate queries in your Performance report, each with its own impressions, and a page that answers a detailed product question may collect several long query rows from one person. To find them, filter Queries with a regex such as ^(how|what|why|which|can|should|is|does). Search Console uses RE2 syntax and partial matching by default. Sort by impressions. Long question queries with impressions but few clicks often mean an AI answer is doing the work on the page.

The generative AI report, and what it leaves out

The Search report and its dimensions

The Search report counts how many times links to your site were shown to a user in a generative AI feature on Google Search ^[2]. It covers AI Overviews and AI Mode, and Google expects to add more. It reports impressions only, and only for the Web search type. You can group by Pages, Countries, Dates and Devices, where Pages means the final URL linked by the AI feature after any redirects. Two results from the same site in one AI feature count as one impression in the chart, and the usual 1,000-row limit applies. It is a visibility report. Clicks live elsewhere.

Discover, multimodal and the data warnings

Discover has its own generative AI report, where an impression counts only once the link is scrolled into view; its help page says both that it reached all sites and that access is still rolling out. Since September 24, 2026, a Web multimodal filter also separates searches from Google Lens, Circle to Search, image uploads and Chrome image search, in both the Search results report and the generative AI reports, though query text is not available for those searches. Google also admits a logging error that reduced generative AI impressions in Search from August 13 to 17, 2026, restored on August 21, plus several Discover logging errors. Check the data anomalies page before reading any AI trend.

What the report cannot tell you

It shows no clicks, no queries and no text of the AI answer. It cannot say whether you were cited first or last, and it cannot show competitors. For clicks, use the Web data, where AI clicks are already counted. The honest way to read AI performance is to set a page's AI impression trend beside that page's ordinary Web clicks over the same dates and look for divergence. Bing goes further, with a citation share metric Google does not yet offer. Paid AI visibility tools only estimate what Google does not publish.

Opting out: the Search generative AI control and its cousins

How the control works

Under Settings, then Search generative AI, owners can choose Include, Exclude or Inherit ^[3]. Include is the default for all properties. Exclude removes the site from AI Overviews, AI Mode and AI features in Discover, including grounding, which makes it the most direct AI switch Google offers. A change generally takes a few days; Google says sites should be excluded within 1-2 days after the control goes live, though caching can delay some content. Child properties inherit from their closest configured parent, and owners of the top domain see a notice when a child overrides it.

What Exclude does not do, and whether to use it

Exclude is not a training opt-out; for that, Google points to Google-Extended. It is not a ranking signal either. Content from other sites still appears in those features, and your content may still power Google Search as a whole. Should you use it? Probably not. Google's own guide says a site must be included in Search generative AI features to be eligible for display in them, so excluding removes a growing surface. Publishers with licensing disputes may decide otherwise. Measure first: a page that is shown often in AI answers and still earns clicks would lose visibility without any obvious gain.

Google-Extended, snippet rules and the robots.txt trap

Google-Extended is a different switch. It controls whether content trains future Gemini models and grounds answers in Gemini Apps and Vertex AI, and its documentation does not mention AI Overviews or AI Mode. Page-level control comes from snippet rules instead. The nosnippet rule applies to all Google search results, including AI Overviews and AI Mode, and stops content being used as a direct input for them. max-snippet caps how much content can feed those features, and data-nosnippet hides specific sections. One trap remains. If a page is blocked in robots.txt, Google never sees its snippet rules, and the page can still be indexed from links. Let Google crawl it and use meta rules.

What Google says actually moves AI visibility

Eligibility is ordinary indexing

For generative AI features, a page must be indexed and eligible to be shown in Google Search with a snippet. Search Console already checks both, with Page indexing for the index and URL Inspection for a single page. Google's AI guide goes further and says that optimizing for generative AI search "is optimizing for the search experience, and thus still SEO". Its AI features are rooted in the core ranking and quality systems. AEO and GEO are labels. Search Console is still the measuring tool, and the reports you already use still matter most.

What you do not need, and what can hurt

Google says you do not need new machine-readable files, AI text files, special markup or Markdown to appear in Search, including its AI features, and that llms.txt will neither harm nor help. There is no requirement to break content into tiny pieces for AI. Structured data is not required for AI features, though it remains useful for rich results. What works, per Google, is unique, useful content with expert or experienced takes. Making a page for every fan-out query variation to manipulate rankings or AI responses breaks Google's scaled content abuse policy. A manual action would follow.

Beyond the click: conversions, Analytics and pages that lose traffic

Google claims that clicks from results pages with AI Overviews are higher quality, with users more likely to spend more time on the site. That is Google's claim, not independent data, and Search Console cannot test it because it does not see time on site. Google Analytics can, using engaged sessions. Google itself recommends looking at sales, signups and a more engaged audience rather than clicks alone. Pair the two tools: Analytics added an AI Assistant channel on May 13, 2026 for visits from ChatGPT, Gemini and Claude. Our Search Console vs Google Analytics comparison explains how the pair splits the work.

If a page holds its AI impressions while its clicks fall, the AI answer may now satisfy the question. Add what an AI summary cannot: data, tools, first-hand detail. Then watch the next 28 days. Other surfaces matter too. Google advises high-quality images and videos for multimodal search, and AI responses can include product listings and local business data drawn from Merchant Center feeds and Business Profiles. Keep those feeds accurate. AI answers inherit their errors.

Connecting Search Console to Claude: no official MCP

What Google does and does not ship

There is no official Search Console MCP server. Scalekit, Quattr and a developer writing on October 4, 2026 all say the same ^[4]. That gap matters because Google already ships official MCP servers for other products, BigQuery, Maps, Drive, Gmail and Google Analytics 4 among them. Anyone connecting Search Console to an assistant today is trusting a third-party author with read access to data Google shows only to verified owners and the users they add. The nearest official route is the BigQuery MCP over your bulk export, where results cap at 3,000 rows and three minutes per call. That needs the bulk export running first.

What an MCP server actually is

A Search Console MCP server lets Claude or another assistant call the Search Console API in conversation. As one guide puts it, it "reads the same data as the API, nothing more". It is a new interface, not new data. The difference from the API is not capability either. A community server exposes only what its author chose to wrap and has no versioning contract, while the API is versioned and supported by Google. For a scheduled pipeline, call the API directly; our Search Console API guide covers it. For questions in a chat, an MCP is faster.

The servers you can use today

Local community servers

AminForou's mcp-gsc is the most popular. It is MIT-licensed, had 1.9k stars on GitHub when we read it, and works with Claude Desktop, Cursor, Codex CLI, Gemini CLI and other MCP clients. Its README counts 20 tools, and destructive tools such as delete_site are off by default. It supports OAuth with your own Google account, which reaches every property you can see, or a service account, which reaches only the properties you add it to as a user. Agencies usually prefer the latter. One limit stands out: the local server works only in the Claude Desktop app, not in the claude.ai browser interface.

ahonn's mcp-server-gsc is the minimal option, also MIT-licensed. It exposes one search_analytics tool with up to 25,000 rows per request, authenticates with a service account only, and includes a quick wins detector for positions 4 to 20 by default. Suganthan Mohanadasan's guide, updated September 27, 2026, describes a server with 29 tools across six categories, while the same page's changelog mentions 20. In Claude Code, setup is one command: claude mcp add gsc -- npx -y suganthan-gsc-mcp. Tool counts vary. Check the repository.

Hosted connectors and SEO suites

Hosted options skip local setup. Ekamoira offers 13 tools with a 30-day trial, Composio's free Hobby plan includes 100,000 tool calls a month, and Windsor.ai offers a free forever plan while noting that Claude Desktop and ChatGPT allow external connectors only on their paid plans. The trade-off is control. A local server keeps credentials on your machine but needs Python or Node.js; a hosted connector stores your Google tokens with a vendor you must trust. Several SEO suites also expose Search Console through their own MCP, among them Ahrefs on paid plans from Lite up and SE Ranking.

Using an MCP server without fooling yourself

Quotas, freshness and scopes

An MCP server is bound by the same API limits: 1,200 Search Analytics queries a minute per site and 2,000 URL inspections a day per site. At that rate, checking a 10,000-page site takes at least five days, so inspect a sample of templates and revenue pages. Freshness is the next trap. The API returns final data by default, which runs 2-3 days behind the dashboard unless you request dataState all; mcp-gsc switched its default to all in version 0.2.0. If Claude's numbers disagree with the dashboard, check this first. Finally, request the read-only scope, webmasters.readonly, unless you need to submit sitemaps or add properties. The API cannot change page content anyway.

Privacy and made-up answers

A local server keeps data off third-party servers, but the rows Claude reads still go to the model provider as part of the conversation. Treat query data as shared with your AI provider. The numbers are real, but the model interprets them, and one author added a verify_claim tool because that reading can go wrong. Good questions match the data: pages that lost clicks month over month, queries on positions 8 to 20 with high impressions, pages with high impressions and low CTR. Bad ones ask for search volume, competitor rankings or backlink totals, which Search Console does not hold. Asking anyway invites made-up numbers.

A weekly AI search routine, with Bing alongside

Bing measures what Google does not. Its AI Performance report, launched in public preview on February 10, 2026, shows how often your content is cited in generative answers across Copilot, along with grounding queries, while Google's report shows impressions. Both are free. A weekly routine needs about twenty minutes: open the generative AI report and note pages with rising or falling AI impressions, compare them with Web clicks for the same pages, check the Search generative AI setting has not changed, then open Bing's report. Keep a simple log of your five pages with the most generative AI impressions and their Web clicks. After two months you will know whether AI features send visitors or answer the question for them.

Who should care most? Publishers, review sites and anyone answering informational queries, because AI Overviews summarize exactly those pages. Sites that rank for brand and transactional queries see little AI impact, and for them the AI report adds one glance a month to the usual Performance report, Page indexing and alerts. Do not rebuild a site for AI search when Google itself says it is still SEO.

The bottom line

Search Console is now the official measuring tool for Google's AI search, with impressions in a dedicated report, clicks folded into Web data and a switch to opt out that most sites should leave alone. Connecting it to Claude works today only through community or vendor MCP servers, which read the same API under the same quotas. Treat them as a faster way to question data you already own, and verify anything surprising in the dashboard.