How to do AIO (GEO) with Claude Code: check, write, publish, check again

You use Claude Code, and you want ChatGPT, Perplexity, Gemini, and Google's AI Overviews to name your brand when a customer asks them for what you sell. This guide sets up that work as a loop you run from Claude Code: ask the AIs your customers' questions, read who they named and which pages they cited, write pages for the questions they answer without you, one per topic, publish them, and ask again once the AI has read it.

Before you set anything up, try one question: the free AI visibility checker asks it on ChatGPT, Perplexity, Gemini, and Google AI Overviews, with no sign-up, and shows whether each one named your brand and linked to your site.

Anthropic's Claude Code docs, and every other vendor page quoted here, were read on November 8, 2026.

What Claude Code can do alone, and what it can't

Claude Code reads and writes files, runs commands, and has two web tools of its own. Per Anthropic's tools reference, WebSearch "performs web searches" and WebFetch "fetches content from a specified URL." With those, it can research a topic, read the pages that rank for it, and write a better one.

It can't see the answer ChatGPT gives your customer. That answer is written inside OpenAI's app at the moment someone asks, and the same goes for Perplexity, Gemini, and AI Overviews. A web search from Claude Code returns search results, not what another company's AI wrote from them. To read those answers, Claude Code needs a tool that asks each AI and hands back what it said: an MCP server, or a script that calls a scraping service with your API key.

The loop in 5 steps

Each step is a sentence you type in Claude Code, or one of the 3 prompts Convika's server adds. Claude Code lists an MCP server's prompts when you type /, shown as /convika:check_ai_answers (MCP), and typing /mcp__convika__check_ai_answers runs it too (Claude Code's MCP docs). Set up the connection first (below).

Save your brand and your customers' questions

Tell Claude Code:

Save these questions in my Convika AI answer check, with "Northside Bakery" as the brand name and northsidebakery.com as my site: "best bakery for a custom birthday cake in Leeds", "where to buy sourdough in Leeds city centre", "gluten-free bakery in Leeds".

Claude Code calls update_answer_check_prompts. Saving asks no AI and uses no checks. Write each question the way a customer types it, without your brand name: a question that names you only tests whether the AI can repeat a name it was given.

Ask the AIs

Run the AI answer check on all my questions, on all 4 AIs. Tell me how many checks it uses before you run it.

Claude Code calls run_answer_check, and Convika asks ChatGPT, Perplexity, Gemini, and Google AI Overviews. The tool tells Claude Code to state the run's cost and the checks you have left, and to run only when you agree. The answers come in over a few minutes.

Read who the AIs named and which pages they cited

/mcp__convika__check_ai_answers

The prompt calls get_answer_check_results and asks no AI. For each question and AI, Claude Code tells you how many answers named your brand ("named in 2 of 4 answers"), whether the newest answer linked to the page meant to answer that question, and the links the answers cited instead. Those links are your reading list: they show the kind of page each AI trusts for the question. To read the same lists by hand, or with other free tools, see how to see which sources ChatGPT cites.

Make pages for the questions the AIs answer without you, and publish them

/mcp__convika__make_pages_for_questions

Claude Code proposes one page per topic, at a path such as /guides/custom-birthday-cakes, and waits for your go-ahead. It then writes the pages, saves them as drafts, points each question at its page, and shows you a private preview link for each. When you've read them, say "Publish them." How the pages are made and published.

Ask again once the AI has read the page

Which of my questions can be asked again since their page was read? Ask only those, on those AIs.

Convika lists the questions whose page ChatGPT's or Perplexity's answer bot has read since it went live, and run_answer_check asks only those. For a page the answers still don't cite, run /mcp__convika__fix_uncited_pages. Checking again after the crawler read.

Set it up

1. Create a free Convika account at app.convika.com. You don't need a credit card.

2. Add Convika's MCP server to Claude Code. Run this in your terminal:

   Copy code

   claude mcp add --transport http convika https://mcp.convika.com
   

   Add --scope user to have it in every project.

3. Sign in. Run /mcp in Claude Code, pick convika, and sign in to Convika in your browser, then click Allow. Anthropic's docs: "Use /mcp to authenticate with remote servers that require OAuth 2.0 authentication." From your shell, claude mcp login convika does the same. Claude Code keeps the sign-in, so you never paste a key into a file or the chat.

4. Pick your questions. Start with the ones a customer asks before buying: "best X in [city]", "X vs Y", "how much does X cost". If you already have live pages on Convika, open AI & search in the dashboard once: it suggests questions from those pages, and saving one costs nothing.

5. Add your main website. Every site your Convika pages are on counts as yours in the check without being typed. If your main site is elsewhere, give its domain when you save the questions, so a link to it counts too.

To check the connection, ask Claude Code "Which Convika workspace are you connected to?". Each of the 3 prompts starts by calling get_brand and stops if the connection works on a workspace other than the one you named.

What the first run costs

• One check is one question answered by one AI. A run of 12 questions on all 4 AIs uses 48 checks. The dashboard shows the same sum before you run: "12 questions × 4 AIs = 48 checks".
• Only an answer that comes back is used. Checks are reserved before anything is asked. One that ends in an error, or comes back empty, such as a search Google shows no AI Overview for, is given back.
• Where the checks come from. Each month, Pro includes 100 and MAX 3,500. They're used first and don't carry over. Free includes none. Packs work on any plan, Free included: from 100 checks for $3, with no expiry. A workspace's first pack, of any size, doubles its base checks.
• Questions you can keep: 10 on Free, 50 on Pro, 200 on MAX.
• One run at a time. While a run is still asking, a new one is refused and uses nothing.
• Claude Code's side. Pages Claude Code writes and saves use none of Convika's AI credits. Claude Code's own work counts toward your Claude plan's usage.

Making the pages, and publishing them

/mcp__convika__make_pages_for_questions takes the questions that have no page yet. Questions that ask the same thing in other words share one page. It shows you the list of paths, page names, and the questions each answers, and does nothing until you agree. Then Claude Code writes each page so it "answers its questions in the first sentences," saves them all as drafts with create_pages, links each question to its page with update_answer_check_prompts, and gives you a preview link per page.

What makes a page one an AI can cite is the same as for search. Google's guide to its AI features says a page must be "indexed and eligible to be shown in Google Search with a snippet," and that "you don't need to create new machine readable files, AI text files, or markup to appear in these features" (Google). So ask Claude Code for:

• the answer in the first 2 or 3 sentences, in the words the question uses
• your real numbers: prices, hours, sizes, delivery areas, with the date you checked them
• a table when the question compares options
• what the pages cited instead cover that yours doesn't

The work off your own site, getting listed on the review sites and lists the AIs cite, is in how to get your SaaS recommended by ChatGPT and Perplexity.

Publishing takes 2 calls. Say "Publish them," and publish_pages first runs as a dry run that lists each page's live URL and any warnings, publishing nothing. The real call needs the token that dry run returned, so Claude Code can't skip it. On Pro and MAX, Convika sends each newly published page to IndexNow, which Bing and other search engines take. Google doesn't take IndexNow. On your own domain, ask Google to index each new page in Search Console. A convika.page address sits in Convika's Search Console, not yours: there, the Google Search tab in AI & search, or get_search_performance over MCP, shows the page once Google has shown it in a search.

Every Convika site also serves /llms.txt, listing its live public pages, and lets AI bots in on its custom domains unless you switch them off. None of this promises a place in an AI's answer.

Checking again after the crawler read

An AI can't cite a page it hasn't read. Convika records verified reads by ChatGPT's and Perplexity's answer bots. Once one has read a page since it was published, get_answer_check_results lists that question and AI under askAgain, and the dashboard shows a button such as Ask ChatGPT again beside the question. Asking again uses checks like any run.

Gemini and AI Overviews read pages as Googlebot, which Convika doesn't record. For those 2, wait until Google has indexed the page before you ask them again. On your own domain, Search Console's URL Inspection tells you. On a convika.page address, wait until the page shows in the Google Search tab in AI & search.

For a page the newest answers still don't cite, run:

/mcp__convika__fix_uncited_pages

Claude Code reads your page, compares it with the pages the answers cited instead, and proposes changes one page at a time. After you agree, it saves each fix as a new version with a preview link. It publishes only when you ask, through the same dry run.

Doing it every week

Claude Code has 3 ways to run a prompt on a schedule. They differ in where they run and whether they can reach a server you added with claude mcp add. From Anthropic's comparison:

Use /loop for the wait after publishing ("check every hour whether ChatGPT has read the new pages"). For the weekly run, use a Desktop scheduled task, or a cloud one with Convika added under Customize → Connectors on claude.ai at https://mcp.convika.com. Anthropic says a cloud task can't see servers you added locally: "MCP servers you added locally in the CLI with claude mcp add are stored on your machine rather than your claude.ai account."

Have the scheduled run stop at drafts:

Read my Convika AI answer check. If the checks I have left cover running every question on all 4 AIs, run it. When the answers are in, list the questions where my brand wasn't named. For the ones with no page, write a draft page each with create_pages and link the question to it. Don't publish. End with the preview links.

You read the previews when you're back and say "Publish them."

If your site isn't on Convika, the free aio-cycle plugin for Claude Code and Codex runs a weekly cycle on any site. It measures the answers of ChatGPT, Perplexity, Gemini, Google AI Mode, and AI Overviews through Scrapeless with your own API key, reads Search Console, and sends new pages to IndexNow. It's MIT-licensed and on GitHub; its scripts need Node.js 20 or newer.

Which plan, and when to pick another tool

• MAX ($49 a month) is for the reader of this guide: a team, or one person, running AIO in-house with Claude Code, Codex, or ChatGPT over MCP. You ask many questions on 4 AIs every week and publish pages for the ones they answer without you, one per topic, on your own domain, at a fixed monthly price with hosting included. MAX includes 3,500 checks a month and keeps 200 questions: enough to ask every one of them on all 4 AIs once a week, though a month with a fifth weekly run needs fewer questions in one run, or a pack. It keeps up to 10,000 pages live, makes up to 1,000 new pages a day, and adds 300 AI credits a month.
• Pro ($12 a month) fits a handful of questions a month: 100 checks a month, 50 questions kept, and 100 live pages, with IndexNow too.
• Packs fit checking once, or now and then, with no subscription: from 100 checks for $3, on any plan, Free included. Other services that sell checks without a subscription: pay-per-check AI visibility tools.

Pick another tool if you need daily tracking across many brands, sentiment, or share-of-voice dashboards, or answers from AIs Convika doesn't ask: Claude, Copilot, and Google's AI Mode. Per answer, MAX's included checks cost less than Peec AI's Starter answers, Ahrefs' AI Visibility Index checks, or Semrush's AI Visibility Base answers, the 3 trackers below that show their prices. Profound shows no price on the pages read, and also writes content with its AI Marketer: cheaper alternatives to Profound. What the trackers add is daily tracking, more AIs, and dashboards. What MAX adds is the rest of the loop at a fixed monthly price you sign up for yourself: making the page, hosting it, and asking again, all from Claude Code. More trackers, by price per check: AI visibility tools compared.

What we saw on convika.com

Convika ran this loop on its own site before the answer check existed, with scripts that are now the aio-cycle plugin. Between October 4 and October 16, 2026, convika.com's Google clicks went from 1 to 20 a day, and its impressions from 59 to 675 a day. The weekly steps, the costs, and the screenshots are in How to get cited by ChatGPT: 20x clicks in 2 weeks.

In the 7 days to November 8, 2026, chatgpt.com was the largest first-touch source of new verified Convika accounts, about a third of them.

The AIs don't name Convika for questions about AIO tools yet. In our run of November 8, 2026, none of 130 answers to 26 such questions, on ChatGPT, Perplexity, Gemini, AI Overviews, and AI Mode, named it. This guide is one of the pages for those questions, and the loop above is how we'll find out whether it works.

Frequently asked questions

Can Claude Code do AIO or GEO by itself?

Partly. Claude Code can research a topic with its web search, write the page, and run your scripts. It can't see what ChatGPT, Perplexity, Gemini, or AI Overviews answer a customer, so on its own it can't tell you which questions you're missing from. Connect an MCP server that asks those AIs, such as Convika's, or run a script with a scraping service's API key, as the aio-cycle plugin does.

Is there an MCP server for AI search visibility?

Yes. Convika's server, at https://mcp.convika.com, asks ChatGPT, Perplexity, Gemini, and Google AI Overviews your saved questions (run_answer_check), returns who was named and which links were cited (get_answer_check_results), and makes and publishes pages on your Convika site (create_pages, publish_pages). It adds 3 prompts for the loop: check_ai_answers, make_pages_for_questions, and fix_uncited_pages. MCP servers for AI search visibility compares it with Peec AI's, Otterly.AI's, Profound's, and the others'.

How many checks does one run use?

One check per question per AI. A run of 12 questions on 4 AIs uses 48 checks, reserved before anything is asked. An answer that errors or comes back empty is given back. Pro includes 100 checks a month, MAX 3,500, and packs start at 100 checks for $3.

Will Claude Code publish pages without asking me?

No. The make_pages_for_questions and fix_uncited_pages prompts tell Claude Code not to publish. publish_pages runs a dry run first and publishes only with the token that dry run returned. Unless your permission settings allow it, Claude Code also asks you before it runs publish_pages, as it does for any MCP tool.

Can Claude Code run the AIO loop every week by itself?

Yes, with a Desktop scheduled task or a cloud scheduled task. A Desktop task runs only while the Claude desktop app is open and the computer is awake. A cloud task runs with your machine off, but reaches Convika only as a connector added on claude.ai, not as a server added with claude mcp add. /loop doesn't fit a weekly job: Anthropic says recurring tasks "expire 7 days after creation".

Does this work in Codex or another MCP client?

Yes. Add https://mcp.convika.com in the client's MCP settings and sign in; the tools are the same. The slash commands in this guide are Claude Code's way of running a server's prompts. In another client, type the same requests as sentences.

How long does it take for ChatGPT to cite a new page?

Nobody can promise a date, and Convika doesn't. What you can see is the step before: when ChatGPT's answer bot has read the page, Convika lists the question under askAgain, and asking again then shows whether the newest answer cites it. On convika.com, Google clicks went from 1 to 20 a day within 2 weeks of starting the loop.

Do I need an llms.txt file for AI answers?

No. Google says you don't need "AI text files" to appear in AI Overviews or AI Mode. Every Convika site serves /llms.txt anyway, listing its live public pages. What a page needs first is to be readable by the AIs' crawlers and indexed.

Does Convika check Claude's or Copilot's answers?

No. Convika's answer check asks ChatGPT, Perplexity, Gemini, and Google AI Overviews. Ahrefs Brand Radar covers Copilot and AI Mode, and Claude on custom prompts. Peec AI tracks Copilot and AI Mode, and both Peec AI and Profound track Claude on their Enterprise plans.

Is MAX worth it if I only want to measure?

Yes, if asking 4 AIs about once a week is enough. Per answer, MAX's included checks cost less than Peec AI's Starter or Ahrefs' AI Visibility Index. Pick a tracker instead for daily tracking, sentiment and share-of-voice dashboards, or AIs Convika doesn't ask, such as Claude and Copilot. MAX pays off most when you also make and host pages for the questions the AIs answer without you, with questions that ask the same thing sharing one page, in one loop from Claude Code.

Sources

• Claude Code: Connect Claude Code to tools via MCP (read November 8, 2026, 01:59 JST)
• Claude Code: Automate work with scheduled tasks (research preview; read November 8, 2026, 01:59 JST)
• Claude Code: Run prompts on a schedule and Schedule recurring tasks in Claude Code Desktop (read November 8, 2026, 01:59 JST)
• Claude Code: Tools reference (read November 8, 2026)
• Claude Code: Configure permissions (read November 8, 2026, 02:48 JST)
• Google Search Central: AI features and your website (last updated December 10, 2025; read November 8, 2026, 02:03 JST)
• Ahrefs Brand Radar (read November 8, 2026, 01:57 JST)
• Peec AI pricing (read November 8, 2026, 01:58 JST)
• Profound pricing (read November 8, 2026, 01:58 JST)
• Semrush: Free AI Visibility Tool (read November 8, 2026, 01:57 JST)
• Semrush AI Visibility pricing (read November 8, 2026, 04:14 JST)
• aio-cycle README on GitHub (read November 8, 2026, 01:59 JST)