Account audits

Fact-check what AI says about your brand: prices, address, hours and services

Pulls what Google's AI answer says for questions about your brand and which sources it relies on, then checks every claim against the current facts on your site and in Business Profile. Finds which source the wrong or outdated detail comes from and plans how to fix that source.

When to use it

Customers quote a wrong price, an old address or a service you no longer offer that they heard from an AI, or you have never checked what AI says about your brand.

What you get

A table labelling every AI claim about your brand as correct, outdated, wrong or unverifiable, the source behind each error and a fix plan split by source type. Nothing changes and no messages are sent.

Does it change anything in my account?

No. The assistant only reads your data and reports back.

Works with

SEO Intelligence, Business Profile

Fill in before sending

[brand name][domain][country][language code][location name or domain, or "none"][paste them, or "none"][competitor][country name in English]

Check whether AI answers give correct information about my brand [brand name]. My site: [domain]. Market: [country], language: [language code]. My Business Profile location: [location name or domain, or "none"]. Answers I copied from ChatGPT or Perplexity: [paste them, or "none"]. This is a read-only, planning job. 1. Prepare the questions. Propose 6 to 8 questions of these kinds for my brand and get my approval: "what is [brand name]", "[brand name] prices", "where is [brand name]" or its address, "[brand name] opening hours", "what services does [brand name] offer", "[brand name] vs [competitor]", "is [brand name] legit". Write them in the market's language. 2. Credit check. Use seo_history to see whether these questions were run recently and call the reusable ones again with the same arguments. 3. Google's AI answers. Run seo_ai_overview for each question: target=[domain], location=[country name in English], language=[language code]. Take the answer text (ai_answer) and the sources. If references is empty, use the links inside ai_answer as the sources. ai_answer is cut at 1,200 characters, so say that claims in the cut part cannot be judged. Record any question with no answer as "no AI answer". 4. ChatGPT and Perplexity. Opus Growth tools do not return the answer text of these engines. If I pasted answers, judge them the same way. If I did not, suggest that I ask these questions in a fresh chat with no history and paste the answers. 5. Collect the correct facts. Read the pricing, contact, about and service pages with site_read_url and text=true (if truncated is true, again with max_bytes=1000000). If I have a location, read the name, address, phone, hours, categories and service list with gbp_location_details. If these two sources disagree, say so separately, because AI cannot know which one to repeat. 6. Compare the claims. Put every concrete claim from each answer (price, package name, address, phone, hours, service, founding year, ownership, comparison) on its own row and give it one of these labels: correct, outdated (was true once), wrong, not on site (cannot be verified). Do not label by guessing. 7. Find the source. For every outdated or wrong claim, open the source the answer cites: read that page with site_read_url and confirm whether the wrong detail is really there. If the source is on my site, say which page and which sentence. If it is a third-party site (a directory, an old news item, a complaints site), give its URL and the page date. Format the output like this. First a four-sentence summary: how many claims are correct, how many are outdated or wrong, and which wrong claim does the most harm. Then a claims table: question, what the AI said, the source it relied on, the correct fact and its source (site page or Business Profile), label. Then a fix plan, each row based on where the error comes from: 1. Source is my site: the page and sentence to change, the old and new text. If an old price or a discontinued service still sits on another page, suggest updating that page or redirecting it with a 301. 2. Source is Business Profile: the field to change, the old and new value. 3. Source is a third party: how to send that site a correction request, with a draft message. Do not send it. 4. The fact is not clearly written anywhere: which page on the site should state it in one clear sentence. Do not change anything and do not send messages. AI answers can take weeks to re-read a corrected source; say so and suggest measuring the same questions again in 4 weeks.

1. Prepare the questions. Propose 6 to 8 questions of these kinds for my brand and get my approval: "what is [brand name]", "[brand name] prices", "where is [brand name]" or its address, "[brand name] opening hours", "what services does [brand name] offer", "[brand name] vs [competitor]", "is [brand name] legit". Write them in the market's language.
2. Credit check. Use seo_history to see whether these questions were run recently and call the reusable ones again with the same arguments.
3. Google's AI answers. Run seo_ai_overview for each question: target=[domain], location=[country name in English], language=[language code]. Take the answer text (ai_answer) and the sources. If references is empty, use the links inside ai_answer as the sources. ai_answer is cut at 1,200 characters, so say that claims in the cut part cannot be judged. Record any question with no answer as "no AI answer".
4. ChatGPT and Perplexity. Opus Growth tools do not return the answer text of these engines. If I pasted answers, judge them the same way. If I did not, suggest that I ask these questions in a fresh chat with no history and paste the answers.
5. Collect the correct facts. Read the pricing, contact, about and service pages with site_read_url and text=true (if truncated is true, again with max_bytes=1000000). If I have a location, read the name, address, phone, hours, categories and service list with gbp_location_details. If these two sources disagree, say so separately, because AI cannot know which one to repeat.
6. Compare the claims. Put every concrete claim from each answer (price, package name, address, phone, hours, service, founding year, ownership, comparison) on its own row and give it one of these labels: correct, outdated (was true once), wrong, not on site (cannot be verified). Do not label by guessing.
7. Find the source. For every outdated or wrong claim, open the source the answer cites: read that page with site_read_url and confirm whether the wrong detail is really there. If the source is on my site, say which page and which sentence. If it is a third-party site (a directory, an old news item, a complaints site), give its URL and the page date.

Format the output like this. First a four-sentence summary: how many claims are correct, how many are outdated or wrong, and which wrong claim does the most harm. Then a claims table: question, what the AI said, the source it relied on, the correct fact and its source (site page or Business Profile), label. Then a fix plan, each row based on where the error comes from:
1. Source is my site: the page and sentence to change, the old and new text. If an old price or a discontinued service still sits on another page, suggest updating that page or redirecting it with a 301.
2. Source is Business Profile: the field to change, the old and new value.
3. Source is a third party: how to send that site a correction request, with a draft message. Do not send it.
4. The fact is not clearly written anywhere: which page on the site should state it in one clear sentence.

Do not change anything and do not send messages. AI answers can take weeks to re-read a corrected source; say so and suggest measuring the same questions again in 4 weeks.

Open in Claude Open in ChatGPT Discuss in the forum

Example result

An example of the output this prompt produces. Its structure comes from the tools' real output; numbers and names are fictional.

Example output (based on a fictional brand, not real customer data)

Summary. Google's AI answer appeared for 6 of the 7 questions and contained 14 concrete claims in total. 9 are correct, 3 are outdated, 1 is wrong and 1 is not on the site. The most harmful is the price: the answer repeats last year's £45 package, and it comes from an old campaign page on your site. Opening hours are written differently on the site and in Business Profile.

Question	What the AI said	Source relied on	Correct fact and source	Label
Pomegranate Pilates prices	"Monthly pass £45"	pomegranatepilates.example/offer-2025/	£65, /prices/	Outdated
Where is Pomegranate Pilates	"Clapham, London"	pomegranatepilates.example/contact/	Clapham, London, /contact/	Correct
Pomegranate Pilates opening hours	"Closed on Sundays"	fitguide.example/pomegranate-pilates	Open Sunday 10:00 to 14:00, Business Profile	Wrong
Does Pomegranate Pilates have reformer	"Group reformer classes"	pomegranatepilates.example/	The site only lists one-to-one reformer	Not on site

Fix plan
1. On your site: /offer-2025/ is still live and says £45. Redirect it to /prices/ with a 301.
2. On your site: add the same opening hours as Business Profile on /contact/ in one line; the site shows no hours today.
3. Third party: the fitguide.example page is dated 2024. A draft correction request is below; it was not sent.
4. Not on site: if you run group reformer classes, say so clearly on the services page; if not, clarify the vague sentence on the homepage.

ChatGPT and Perplexity answers were not pasted, so only Google was assessed. Measure the same 7 questions again in 4 weeks.

More prompts

Account auditsRead-only

Full Google Ads account audit with the three fixes to make first

Reviews the whole account in one pass, from measurement and bidding to structure and landing pages. Ranks the findings by impact and names the first three fixes without changing anything.

When to use itWhen you inherit or doubt a Google Ads account and want to know what is actually wrong and where to start fixing it.

Audit my Google Ads account [account name] end to end. If the account is younger than 90 days use every day available, otherwise use the last 90 days. This is a read-only review, do not change anything.

Start with account_snapshot to get each campaign's status, budget, bidding strategy, age and results since it started. Separate the campaigns that bring results from the ones that spend without converting. Then work through these areas in order:

1. Measurement. List the primary conversion actions with list_conversion_actions. Flag any primary action that recorded nothing in the last 30 days. Look for setups that can count the same person twice, for example a phone click and a WhatsApp click on the site set up as two separate primary actions, or the Google Ads tag and an imported GA4 event both primary for the same form. Check how many clicks went to the call button in the ads and whether those calls were actually recorded, using google_ads_query on call_view.
2. Bidding and budget. Use impression_share_analysis to show, per campaign, the impression share lost to budget and to rank. Use keyword_quality_report to find keywords below the first page bid or rarely shown because of low quality. For campaigns on Smart Bidding, say whether the last 30 days hold enough conversions for the target they use.
3. Structure. Check whether brand searches share a campaign with generic searches, whether every ad group sends people to the page for its own service, and whether the same keyword sits in more than one ad group.
4. Search terms. Use search_terms_report to separate terms that spend without converting, competitor brands and informational searches. Never count a term with at least one conversion as waste.
5. Ads and assets. Review ad strength, disapproved assets and whether sitelinks point to the right page. Flag separately any copy or asset that describes a service the business does not offer: [services not offered, or write "none"].
6. Landing pages. Pull clicks, cost and conversion rate per page from landing_page_view with google_ads_query. On the pages that take the most spend, check that a phone and a WhatsApp button exist and check mobile speed with core_web_vitals.
7. Location, schedule and device. Show whether the location option is set to "Presence", how much spend goes outside the target area, and the spend and conversions outside [business hours].

Format the output like this: first a five-sentence summary of where the account stands. Then a findings table ranked by importance: problem, evidence, estimated impact, recommendation. At the end, the three fixes to make first and why each one comes first. Where the data is not enough for a conclusion, do not guess, say what is missing.

Account auditsRead-only

Full Meta Ads account audit with the three fixes to make first

Reviews measurement, delivery issues, the learning phase, targeting, placements and creative fatigue in one pass. Ranks the findings and names the first three fixes without changing anything.

When to use itWhen your Meta ads results have slipped or stalled and you need to know what is wrong in the account before touching anything.

Audit my Meta ad account [ad account] end to end. Use the last 30 days as the main period and the 30 days before that as the comparison. This is a read-only review, do not change anything.

Start with account_snapshot for platform meta to get each campaign's status, budget, objective, age and results since it started. Meta has no single conversion count, so use the result each campaign actually optimizes for (leads, messaging conversations started, purchases). Then work through these areas in order:

1. Measurement. Check the pixel's last event time with list_meta_pixels. Confirm that the event each campaign optimizes for is actually recording. Flag setups that can count the same result twice, for example leads collected by both the pixel and an instant form.
2. Delivery. List ad sets that are rejected, stuck in review, have no active ad left, or spend less than half of their budget. Give Meta's stated reason for each.
3. Learning phase. Find ad sets split so thin that they cannot reach 50 results a week. Use meta_change_history to show edits in the last 14 days that reset learning.
4. Targeting and audiences. Flag ad sets that target the same audience and ad sets whose frequency is above [maximum frequency].
5. Placements and devices. With get_report, compare cost per result by publisher_platform and by device against the account average. Do not decide on segments with fewer than 20 results.
6. Creatives. Treat an ad as fatigued when frequency rises while click-through rate falls and cost per result climbs. Briefly describe the angle the winning ads share.
7. Destination. Check that the page or WhatsApp link the ads send people to works and matches the objective.

Account auditsChanges after your approval

Merge phone and WhatsApp enquiries into one reliable conversion

Stops one person being counted twice when they both call and message. Ties the two buttons on the site to a single conversion and checks whether the call button in the ads is measured at all.

When to use itWhen people call and message you from the same site and you suspect one visitor is being counted as two conversions, making your cost per conversion look better than it is.

In my Google Ads account [account name], prospects reach us through the phone and WhatsApp buttons on the site. If one person pressing both buttons counts as two conversions, my cost per conversion looks cheaper than it is and the bidding algorithm learns the wrong thing. Fix this.

1. Map the current setup. List the primary actions, their categories and counting types with list_conversion_actions. In my Tag Manager container [GTM container ID], use list_gtm_tags to show which tag fires which conversion label. Estimate with numbers how often one ad click may have produced more than one contact conversion in the last 30 days.
2. The call button in the ads. With google_ads_query, check how many clicks and how much spend went to the call button in the ads, and use call_view to check whether those calls were recorded. If there are clicks but no records, say so clearly and offer two options: remove the call asset, or accept those clicks as unmeasured traffic.
3. Show the plan. Propose a single conversion action named "Contact click (Phone + WhatsApp)" with counting set to one. Tie it to one tag that fires on both the phone and the WhatsApp trigger. Make the two old actions secondary. Say whether campaigns on Smart Bidding need their target reviewed, because the conversion count will drop.
4. Apply after my approval. Create the conversion action, set counting to one, update the tag in Tag Manager and show me the list of changes before publishing. After publishing, write down which version can be restored.

Finish with a short test plan: how to see in preview mode that only the new tag fires when both buttons are pressed, and how many days it takes for the fix to show in real data.

Account auditsRead-only

Run a full SEO, AEO and GEO audit and get an evidence-backed scorecard

Scores access, technical health, structured data, quotability of content, trust signals and measured AI visibility in six areas of five concrete checks each. Shows the evidence for every point and orders the fixes by impact and effort, without changing anything.

When to use itYou want one evidence-backed report showing where your site stands in Google and in AI search, and a clear task list to hand to an agency or developer.

Audit my site [site URL] end to end for classic search, answer engines (AEO) and AI search (GEO). My Search Console property: [Search Console property]. Main competitor: [competitor domain, or "none"]. Market: [country], language: [language code]. The 3 searches that bring me the most business: [3 commercial queries]. My most important page templates: [homepage, one service page, one blog post URL]. This is read-only.

Credit rule: before starting, check earlier runs with seo_history and call the reusable ones again with the same arguments. Show me once the list of new calls that will spend credits and get my approval.

Data collection:
1. Run the bundled analysis with seo_full_analysis, passing domain, competitor, location and language. Its AI visibility section measures the single keyword the site ranks for most, which may have nothing to do with my business; do not use that section for scoring, and instead measure my 3 commercial queries with seo_ai_overview and target=[domain]. If references is empty, count the domains of the links inside ai_answer.
2. Run seo_site_crawl with max_pages close to the site's page count (at most [page limit]). It returns a task_id; call it again with the same task_id after 60 to 90 seconds and wait until the summary arrives. Pull the full URL list of an issue with the issue parameter when needed. If the crawl limit does not cover the whole site, say the results are a sample.
3. Read robots.txt, sitemap.xml and llms.txt with site_read_url. Read my template pages with text=true (if truncated is true, again with max_bytes=1000000) and with outline=true. For JSON-LD, read the same pages with max_bytes=1000000 and extract the blocks.
4. Measure the homepage and the service page with seo_audit and strategy='mobile'. For Core Web Vitals use real-user (field) data where it exists, otherwise the lab value, and say which one you used.
5. Use get_report with platform='search_console' and breakdown 'query by impressions' and 'page by impressions' for the last 90 days. Inspect the 10 pages with the most impressions with gsc_inspect_urls. Get the sitemap status with gsc_sitemaps.
6. Remember that different tools can see different versions of the same page: if seo_technical_audit or seo_site_crawl measures a title in a different language from what site_read_url shows, record it as a finding, because the site may be redirecting visitors by country.

Scorecard: six areas, five checks each, each check pass (1) or fail (0). If a check has no data, write "not measured" and drop it from the denominator; never score by guessing.
1. Access and crawlability: robots.txt is open to search-role AI crawlers; key pages return 200 with at most one redirect; no noindex on key pages; the sitemap has no errors and contains the key pages; the main content is visible without JavaScript.
2. Technical health and speed: mobile LCP under 2.5 seconds; CLS under 0.1; INP under 200 ms; no broken pages or broken internal links in the crawl; no duplicate titles or descriptions.
3. Structured data and entity: an Organization or LocalBusiness node exists; nodes are linked by @id; the schema fits the page type (Service, Article, Product); schema facts match the visible text; sameAs URLs work.
4. Quotability of content: the first paragraph of key pages answers the question directly; headings name the topic or question clearly; comparable facts are in a table; a page-specific FAQ exists; there are concrete numbers and dates.
5. Trust signals: the about page is concrete (who, where, since when); contact details are on the site and consistent; posts show an author or publisher; publish and update dates are visible; legal pages are reachable from the footer through static links.
6. Measured AI visibility: cited in Google's AI answer for at least one of the 3 commercial queries; cited for the brand query; llms.txt exists and is accurate; Search Console shows clicks from non-brand queries; if the main competitor is cited on these queries, there is at least one where I am cited too.

Format the output like this. First a five-sentence executive summary. Then the score table: area, score (e.g. 3/5), failed checks. Then the evidence table: area, check, result, evidence (tool name and returned value), affected URLs. Then the fix list: for each item the impact (high, medium, low), the effort (hours, days, weeks) and who does it; high impact and low effort first. At the end, the first three fixes and why they come first. Do not pass judgement on anything that was not measured.