Is ChatGPT Recommending You? Track AI Mentions with Claude
Write 20 prompts a buyer would ask ("best CRM for a 10-person agency"), run them through Apify's official ChatGPT, Perplexity and Google AI Overviews scrapers, and have Claude count which brands each answer names and which domains it cites. Run it weekly as a Claude Cowork scheduled task. On Apify's Starter plan, 20 prompts across ChatGPT, Perplexity and AI Overviews cost about $0.40 a week. On the Free plan the ChatGPT and Perplexity scrapers charge $0.20 per result, so test small there.
A growing share of buyers ask an AI assistant before they ever see a search results page. If ChatGPT lists four tools for "best invoicing software for freelancers" and yours is not one of them, no rank tracker will tell you. This post builds the check yourself: a fixed prompt set, the AI answers scraped every week, and Claude turning them into a share-of-voice table you can compare week to week.
What you are measuring
Two different things, and it pays to keep them apart:
- Mentioned. Your brand name appears in the answer text. This is what a reader sees, and it is the closer proxy for "recommended".
- Cited. Your domain appears in the answer's sources. AI search engines often cite a page for a fact while recommending a competitor in the prose, so a citation is not a recommendation.
A brand that is cited but never mentioned has content the engine trusts but a product it does not suggest. A brand that is mentioned but never cited is being recommended on the strength of other people's pages, usually reviews, roundups and Reddit threads. The two numbers point to different fixes, which is why the table at the end tracks both.
Let Claude set it up for you
Paste this into Claude Cowork and it walks you through the whole build: it asks for your brand and prompts, checks the Apify connector, runs a three-prompt test with the cost shown, then creates the weekly task. The step-by-step sections below explain what each part does.
Claude interviews you, tests on three prompts, then schedules a weekly share-of-voice report.
- Asks for your brand, competitors, market and prompts
- Checks the Apify connector is connected
- Runs a 3-prompt test and shows the cost
- Creates the weekly scheduled task after you confirm
Help me set up a weekly check of whether AI search engines mention my brand.
Work through this one step at a time and ask me one question at a time.
1. Ask me, in turn: my brand name and domain; 3 to 6 competitor names and
domains; my main market as a two-letter country code; which engines to
track (offer ChatGPT, Perplexity, Google AI Overviews, Google AI Mode,
Gemini, Microsoft Copilot); and my buyer prompts. If I have no prompts,
draft 20 for my category (8 category, 6 problem, 4 comparison, 2 brand)
and let me edit them. Then repeat the frozen list back to me.
2. Check that the Apify MCP connector is available by calling
fetch-actor-details on apify/chatgpt-search-scraper. If it is not
connected, stop and tell me to add a custom connector in Customize >
Connectors with the URL
https://mcp.apify.com?tools=apify/chatgpt-search-scraper,apify/perplexity-search-scraper,apify/google-ai-overviews-scraper,apify/google-ai-mode-scraper,apify/gemini-search-scraper,apify/microsoft-copilot-search-scraper,runs,storage
then wait for me to confirm.
3. Use only these Actors, for the engines I chose:
apify/chatgpt-search-scraper, apify/perplexity-search-scraper,
apify/google-ai-overviews-scraper, apify/google-ai-mode-scraper,
apify/gemini-search-scraper, apify/microsoft-copilot-search-scraper.
Input: "queries" (one prompt per line) and "countryCode" (not
supported by the AI Overviews Actor).
4. Before any run, read each chosen Actor's per-result price with
fetch-actor-details, estimate the cost of a 3-prompt test and of the
full weekly run, and ask me to confirm. Run the test only after I say yes.
5. From the test results (fields: query, text, sources), show for each
prompt and engine which brands are named in the text, in what order,
and whether each brand's domain is in sources. Count explicit brand
names only, never inferred ones. If an engine returned no results,
say so and leave it out of the weekly task unless I ask otherwise.
6. If I am happy with the test, show me the weekly task you will create:
name, schedule (every Monday), and full instructions. The instructions
must run the full prompt list, build a share-of-voice table and a
citations table (brand x engine), list prompts where a competitor is
named and I am not with the sources cited, then use get-actor-run-list
and get-dataset-items to compare against last week's runs and end with
one CSV row per brand. Create the scheduled task only after I confirm.
Claude asks before each change it makes. Read each step before you approve it.
Once the task is running, two follow-ups are worth keeping to hand.
From this week's results, take every prompt where a competitor is named and
we are not. Group the cited source URLs by domain, rank the domains by how
many lost prompts cite them, and tell me which are reviews, roundups, forums
or vendor pages.
Run this one prompt three times on each engine I track and tell me in how
many of the three answers each brand is named: [prompt]. Show the cost
before you run it.
The Actors, one per AI surface
Apify maintains its own scraper for each major AI answer engine. All six take the same core input, a queries field with one prompt per line, and all are pay-per-event, so there is no separate compute charge on top. Prices below are per result, read from each Actor's pricing data on 2026-10-21. The Starter prices for ChatGPT, Perplexity, Gemini and AI Overviews match what a Starter account was actually charged in a test run that day.
| AI surface | Actor | Free plan | Starter plan |
|---|---|---|---|
| ChatGPT Search | ChatGPT Search Scraper | $0.20 | $0.005 |
| Perplexity | Perplexity Search Scraper | $0.20 | $0.013 |
| Google AI Overviews | Google AI Overviews Scraper | $0.003 | $0.002 |
| Google AI Mode | Google AI Mode Scraper | $0.20 | $0.005 |
| Gemini | Gemini Search Scraper | $0.20 | $0.005 |
| Microsoft Copilot | Microsoft Copilot Search Scraper | $0.20 | $0.005 |
Each also charges an Actor-start event of $0.00005, which rounds to nothing.
In that same test, two prompts per engine, ChatGPT, Perplexity, Gemini and AI Overviews each returned both answers. Google AI Mode and Microsoft Copilot returned none: the AI Mode run timed out waiting for its upstream provider, and Copilot reported "response contained no answer text" for both prompts. Neither run charged for the missing results, only the $0.00005 start fee. One failed test does not mean these two Actors are broken, but run a small test of each before you add it to a weekly task.
The Free plan column is the detail most guides skip. Five of the six Actors charge 40 times more per result on Free than on Starter. Twenty ChatGPT prompts cost $4.00 on Free, most of the $5 monthly credit in a single run, against $0.10 on Starter. Use Free to see the output shape with two or three prompts, and treat Starter as the real floor for a weekly tracker. Google AI Overviews is the exception, cheap on both plans, so it is the one surface you can track properly without paying for a plan.
Inputs beyond queries, from each Actor's input schema:
countryCode(all except AI Overviews): the country the prompt is sent from. It defaults to the United States. Set it to your market, because answers differ by country.perplexitySearch(Perplexity): toggles for returning images and related questions. Leave both off for this job.copilotSearch.copilotSearchMode(Copilot):chat,reasoning,smart(the default) orstudy.
The ChatGPT Search Scraper's dataset has three fields per prompt: query, text (the answer) and sources, a list of {title, url, section} objects. The AI Overviews Actor adds a type field. That shape is everything Claude needs.
Step 1: Write the prompt set
The prompt set is the whole method. Get it wrong and you are measuring nothing.
Write 20 prompts in the words a buyer uses, not the words your marketing team uses. A useful split:
- 8 category prompts: "best project management tool for a remote design team"
- 6 problem prompts: "how do I stop clients paying invoices late"
- 4 comparison prompts: "Asana vs Monday for a 15-person agency"
- 2 brand prompts: "is [your brand] any good", "[your brand] alternatives"
Freeze the list. Changing the wording changes the answers, so a prompt you edit in week 6 breaks the comparison with weeks 1 to 5. Add new prompts as new rows rather than rewriting old ones.
Also decide your competitor list now, 3 to 6 names, so Claude counts the same brands every week.
Step 2: Connect Apify to Claude with only the tools this job needs
Follow the Cowork connector setup: add a custom connector pointing at Apify's remote MCP server. For this job, put the Actors in the connector URL so Claude does not have to search the whole Store every run. Apify's MCP documentation describes a tools URL parameter that accepts Actor names and tool categories:
https://mcp.apify.com?tools=apify/chatgpt-search-scraper,apify/perplexity-search-scraper,apify/google-ai-overviews-scraper,runs,storage
The runs and storage categories add get-actor-run-list and get-dataset-items, which the weekly comparison in step 4 relies on. Add the AI Mode, Gemini or Copilot Actors the same way if those surfaces matter in your market. The same URL works in Claude Desktop and Claude Code; the Claude Desktop guide covers the config file route.
Step 3: Run the first audit
Start with one manual run, in any Claude client, before you schedule anything:
Run apify/chatgpt-search-scraper, apify/perplexity-search-scraper and
apify/google-ai-overviews-scraper with these 20 queries, countryCode "gb"
where supported:
[paste your 20 prompts, one per line]
Then read every answer's text and sources. For each prompt and each engine,
record:
- which of these brands are named in the answer text: [you], [competitor A],
[competitor B], [competitor C]
- the position of each brand in the answer (1 = first named)
- whether each brand's domain appears in sources
Output:
1. A share-of-voice table: brand x engine, % of prompts where the brand is named
2. A citations table: brand x engine, % of prompts where its domain is cited
3. The prompts where a competitor is named and we are not, with the
sources that answer cited
Count only explicit brand names. Do not infer a mention from a description.
The last line matters. Without it, a model can count "a popular all-in-one CRM" as a mention of whoever it thinks that is.
Here is real output from a test run on 2026-10-21, trimmed to one prompt, "best CRM for a small marketing agency", sent from the United States:
| Engine | First brands named in the answer | Sources returned | First cited domains |
|---|---|---|---|
| ChatGPT | HubSpot, Pipedrive, GoHighLevel | 8 | management.org, closefa.st, toolradar.com |
| Perplexity | Pipedrive, HubSpot, Capsule | 10 | nethunt.com, dupple.com, insightly.com |
| Gemini | HubSpot, GoHighLevel, Pipedrive | 2 | nimble.com, pcmag.com |
| Google AI Overviews | HubSpot, Zoho, ActiveCampaign | 5 | reddit.com, nethunt.com, taskip.net |
Even two prompts showed the mentioned-versus-cited gap. Perplexity's first source was NetHunt's own "best CRM for marketing agency" article, yet NetHunt is not named anywhere in Perplexity's answer. Google's AI Overview did name NetHunt, from the same article. One page, two engines, opposite outcomes.
Two smaller things the raw output shows. ChatGPT's source URLs carry utm_source=chatgpt.com, so ChatGPT referrals can be identified in your analytics. And on the second test prompt, "best invoicing software for freelancers", Google's AI Overview named no product at all and pointed to NerdWallet's guide instead. On that prompt, the page to be on is NerdWallet's.
Run your own prompt set across all your competitors and the same table becomes a share-of-voice percentage per brand per engine.
The third output is the useful one. A list of "prompts we lose, and the pages the engine cited instead" is a to-do list: those cited pages are the reviews, roundups and threads you need to be on.
Step 4: Make it weekly
Once the manual run reads well, turn it into a Cowork scheduled task. Scheduled tasks run weekly (or hourly, daily, on weekdays), keep running when your computer is asleep, and can use your connectors, per Anthropic's help article. In Cowork, open Scheduled, click New task, and either let Claude draft it or set it up manually with this prompt:
Every Monday: run the same three Apify Actors on the 20 queries below
[paste the frozen list]. Build the share-of-voice and citations tables
exactly as before.
Then use get-actor-run-list to find last week's run of each Actor and
get-dataset-items to read its results. Show the change in each brand's
share of voice since last week, and list any prompt where we gained or
lost a mention.
Comparing against last week's Apify runs rather than a file keeps the history where the data already lives. Unnamed datasets do expire, though: on the Free plan only the 10 most recent runs are kept, and three Actors a week fills that in about three weeks. A week-over-week delta is always within reach. For a 6-month trend line, ask the task to end with one CSV row per brand and paste it into a sheet you own.
Reading the numbers honestly
Three caveats belong in any report you send upward.
AI answers are not stable. The same prompt can name different brands an hour apart. One sample per prompt per week is a noisy measurement, so read trends over several weeks, not one week's swing. If a single prompt matters a lot, run it two or three times and count how often you appear.
A scraper is not your buyer. These Actors send each prompt fresh, from the country you set, with no chat history or memory. A signed-in user who has been asking about your category for a month may see something different. Treat the numbers as a consistent baseline, not as what every user sees.
Check the terms. These Actors query AI products on your behalf. Each AI vendor sets its own rules on automated access, and they do not all agree with their own robots.txt. Read which AI sites you can scrape before you scale this past a weekly check.
When to use a dedicated tool instead
If you need hundreds of prompts, many markets, dashboards for a team, and history someone else maintains, a dedicated AI-visibility platform (Profound, Semrush's AI toolkit and Ahrefs Brand Radar all sell this) will be less work than a scheduled Claude task. The trade-off is that they choose the prompts and the method, and the price is a subscription.
The setup here fits the other case: you want to own the prompt list, see the raw answers behind every number, and spend a few dollars a month rather than a few hundred. It also lives next to the rest of your SEO data. The same Apify account runs the Google SERP scrapers and the SEO workflows in this guide, so you can put classic rankings and AI mentions in one report.
FAQ
Write a fixed set of prompts a buyer would ask, run them through Apify's ChatGPT Search Scraper, and have Claude count how often your brand is named in the answer text and how often your domain appears in the cited sources. Repeat weekly with the same prompts so the numbers are comparable.
On Apify's Starter plan, per-result prices checked on 2026-10-21 are $0.005 for ChatGPT, $0.013 for Perplexity and $0.002 for Google AI Overviews, so 20 prompts across all three cost about $0.40 a week. On the Free plan the ChatGPT and Perplexity scrapers charge $0.20 per result, which makes a weekly tracker impractical there.
Mentioned means your brand name appears in the answer text the user reads. Cited means your domain appears in the answer's sources. Engines often cite a page for a fact while recommending a different brand, so track both separately.
AI answers are generated fresh each time, so the same prompt can name different brands on different runs. Read trends across several weeks, and run your most important prompts more than once per check.
Yes. Cowork scheduled tasks can run weekly, use your connected Apify MCP server, and keep running while your computer is asleep. The task can read last week's Apify runs to report changes in share of voice.
Next steps
Connect Apify to Claude using the Cowork setup guide, write your 20 prompts, and run the audit once by hand before you schedule it. For brand mentions on Reddit rather than in AI answers, see Reddit brand monitoring in Claude Cowork; to write pages that answer those buyer questions, start from a Claude Code content brief built from the live top 10. More marketing prompts are in Claude prompts that use live web data.
Create an Apify account and run your first AI-search audit →