Key takeaways
- A wrong answer on a direct brand query does more damage than being missing from a niche category search, so audit accuracy first and visibility second.
- You need about 30 minutes, a spreadsheet, and the habit of running every prompt several times, because AI answers change from run to run.
- Most errors trace back to your own product, pricing and landing pages, or to a short list of third-party review and comparison sites.
- Score each answer from 1 to 5, log the likely source of every error, and fix the source page rather than waiting for the model to forget.
- Once the manual audit works, a monitoring platform can repeat it across hundreds of prompts and markets.
Why accuracy deserves its own audit
Most AI visibility advice is about whether you show up. That matters, but it skips a nastier problem: you show up and the answer is wrong. Old pricing. A feature you dropped two years ago. A competitor's certification credited to you. Or your brand mixed up with one that has a similar name.
Buyers read these answers and rarely click through to check them. So a short audit that looks only at correctness is worth doing before any bigger GEO project.
A quick note on what the answers are built from. Promptwatch's data on ChatGPT citations (Sep 28 to Oct 4, 2026) shows that for prompts naming a specific company, product pages make up 23.3% of citations and landing pages 22.8%. Almost half of what ChatGPT quotes about a named brand is your own site. That is good news, because it means many errors are fixable by you.
The flip side: for brand-specific prompts, ChatGPT's source list is short. Five domains is the most common count, and 40.7% of responses cite five or fewer, against 47.6% of all responses citing ten or more. With so few sources, one bad review page or stale forum thread carries a lot of weight.
What you need before you start
- 30 minutes of uninterrupted time
- A spreadsheet with these columns: prompt, platform, run number, score (1 to 5), what was wrong, cited sources, suspected cause
- Accounts or access for ChatGPT, Gemini, Perplexity and Google (for AI Overviews). Claude and Copilot are worth adding if your buyers use them.
- A one-page "source of truth" about your brand: legal name, category, pricing, key features, certifications, founding facts, who you serve, and what you do not do
- A logged-out or private browser window, so personalization does not skew the results
That source of truth document is the step people skip. You cannot grade an answer as wrong if you have not written down what right looks like.
The 30-minute plan at a glance
| Minutes | Step | Output |
|---|---|---|
| 0 to 5 | Write the source of truth and pick 8 to 10 prompts | Fact sheet and prompt list |
| 5 to 15 | Run prompts across platforms, 2 to 3 times each | Raw answers and screenshots |
| 15 to 20 | Score every answer 1 to 5 | Scored spreadsheet |
| 20 to 27 | Trace each error to a source | Cause column filled in |
| 27 to 30 | Prioritize fixes | Top three actions |
Step 1: build your prompt list (5 minutes)
Avoid the mistake of only asking an AI about yourself. Branded prompts tell you what the model says when it is already thinking about you. Category prompts tell you what it says when it is not. You need both, but they test different things.
Pick eight to ten prompts across these types:
- Direct brand questions: "What is [Brand]?" and "Who is [Brand] for?"
- Fact checks: "How much does [Brand] cost?" and "Does [Brand] have [certification or feature]?"
- Reputation questions: "Is [Brand] any good?" and "What are the downsides of [Brand]?"
- Comparison prompts: "[Brand] vs [Competitor]"
- Category prompts: "Best [category] for [your audience]"
Write them the way a buyer would, not the way your marketing team would. If you serve distinct audiences, add one persona-flavored prompt for each, such as "best [category] for a 10-person agency."
Step 2: run the prompts and capture everything (10 minutes)
Run each prompt in ChatGPT, Gemini and Perplexity, and check Google for an AI Overview on the same query. Then run it again. Then once more if time allows.
This repetition is not optional. AI responses are non-deterministic, so a single clean answer proves little and a single wrong answer might be a fluke. You are looking for patterns: the error that shows up in two of three runs is the one to worry about.
For each run, do three things:
- Screenshot the full answer, including the sources list.
- Paste the key claims about your brand into the spreadsheet.
- Copy the cited domains into the sources column.
Do not try to analyze while you capture. Collect first. It is faster, and you will see patterns more clearly when you read the sheet afterward.
If you want a second framework to compare against, PartnerStack's 30-minute audit guide splits the time similarly: platform testing first, then a review of your highest-priority owned pages.
Step 3: score what you find (5 minutes)
Use a simple 1 to 5 rubric. It is adapted from the scale CueForGood published for its own 30-minute audit, and it works because it punishes material errors harder than omissions.
| Score | What it means |
|---|---|
| 5 | Brand name, category, differentiators and values all correct |
| 4 | Core info is right, but certifications, sourcing story or impact claims are missing or vague |
| 3 | Basic category info present, but significant errors or omissions |
| 2 | Material errors: wrong description, conflated with another brand, misleading |
| 1 | Absent, misrepresented, or associated with negative attributes |
Also mark each answer with one of three labels for accuracy: correct, partial, or incorrect. The numeric score helps you track change over time. The label helps you sort fast.
A few judgment calls that come up:
- A missing differentiator is a 4, not a 2. Omission is annoying, not damaging.
- Wrong pricing is at least a 3, and a 2 if it would change a buying decision.
- Being merged with a similarly named company is a 2 even if the rest is right.
- If the answer is accurate but lukewarm, note that separately. Tone is a reputation question, not an accuracy one.
Average your scores by platform and by prompt type. Many brands find that Perplexity, which leans on live sources, is more accurate than a model answering from training memory. If you see that gap, it tells you where to look next.
Step 4: trace every error to a source (7 minutes)
This is where the audit pays off. For each error, ask where the wrong claim could have come from. There are usually four answers.
Your own pages say it, or fail to say it
Open the cited page. Is the old price still on it? Does the product page describe a feature set you retired? Is there simply no page that states the fact?
This is the most common and most fixable cause. Given that product and landing pages dominate citations for brand prompts, an outdated pricing page is a prime suspect. It also lingers: in Promptwatch's half-life study, which tracked pages ChatGPT cited on July 6, 2026 for 90 days, 78% of landing pages and 76% of docs were still cited after 30 days. A wrong fact on one of those pages does not fade quickly. Fix the page; do not wait.
A third-party page says it
Look at the non-owned domains in the sources column. Review sites, comparison articles, old forum threads and directory listings all shape the description. If a comparison page lists you as "missing" a feature you added last year, that page may be what the model repeated. Promptwatch's advice here is to check that third-party comparison and review pages describe you accurately, and to publish your own comparison pages so there is a version you control.
Nothing says it, so the model guessed
If the error has no obvious external source, it is likely a hallucination: the model filling a knowledge gap with a plausible guess. The cure is not a correction request. It is more authoritative, consistent content about the fact across credible places, so there is something real to retrieve.
Old training data
Some claims come from what a model learned before you changed. These are the hardest to move directly. Fresh, crawlable, well-cited pages are what shift retrieval-based answers, and over time the models catch up.
Write the suspected cause next to each error. Do not agonize over certainty. A best guess is enough to pick the right fix.
Red flags that need attention today
Some results should jump the queue:
- Wrong pricing or plan limits
- Claims of certifications, compliance or safety you do not hold, or denial of ones you do
- Confusion with another company
- Negative attributes with no source behind them
- Competitors recommended for something you clearly do
These hit buyers at the moment of decision. Everything else can wait for the next sprint.
Step 5: turn results into a fix list (3 minutes)
You should now have a sheet with scores, errors and suspected causes. Sort by two things: how bad the error is and how many runs repeated it. Take the top three and write the action next to each.
| Cause | Typical fix |
|---|---|
| Outdated own page | Update the page, add a visible last-updated date, and make sure pricing and features are in plain text rather than images or widgets |
| Missing own page | Publish a plain-language FAQ or fact page stating the claim directly |
| Wrong third-party review or comparison | Contact the publisher with the correct facts, update your profile on the review site, and publish your own comparison page |
| Hallucination | Add the fact to your homepage, product pages and docs, and seed consistent statements on credible third-party sites |
| Brand confusion | Use clear entity signals: consistent name, category, location and schema markup across your site and profiles |
Two practical notes. First, check that AI crawlers can actually read the pages you are fixing. A corrected page nobody can fetch fixes nothing. Free tools like the AI crawler URL inspector and the robots.txt generator for AI make this quick. Second, put a date on the fix and re-run the same prompts in two to four weeks. That before-and-after is the only honest proof it worked.
What the manual audit cannot do
Be realistic about the limits. Ten prompts across three platforms is a spot check. It will catch glaring problems and teach you how your brand is described. It will not tell you how the answer varies by country, language, persona or phrasing, and it will not track change week to week without you repeating the work.
The pattern in Promptwatch's research is that one prompt tells you little and you should measure across many. When you outgrow the spreadsheet, a monitoring platform does the repetition. Tools like Promptwatch pull the individual claims AI makes about your brand into one view (its Brand Narrative feature), let you flag a claim as inaccurate, and point to the pages on your site that state it, which maps neatly onto steps 3 and 4 above.
There are plenty of other options, and the right one depends on budget and scale:
| Tool | Angle | Good fit |
|---|---|---|
| Promptwatch | Monitoring plus claim-level accuracy checks, crawler logs, and content creation to fix gaps | Teams who want to find errors and fix them in one place |
| Otterly.AI | Affordable brand monitoring | Small teams starting out |
| Peec AI | Visibility tracking with suggestions | Marketing teams wanting a lighter dashboard |
| Waikay | Brand perception and visibility tracking | Teams focused on how AI describes them |
| Profound | Enterprise visibility analytics | Large brands with bigger budgets |
If you want to browse the wider field, the GEO software directory lists far more options than fit here.
A repeatable cadence
Do not make this a one-off. A sensible rhythm:
- Monthly: rerun the same ten prompts and compare scores to the last run
- After any pricing, product or positioning change: run the fact-check prompts within a week
- After a rebrand, acquisition or launch: run a full audit, including the confusion checks
- Quarterly: refresh the prompt list with new competitors and new buyer phrasing
Keep every spreadsheet. A year of scored audits shows which fixes moved the needle and which were busywork.
Common mistakes
- Testing only your brand name. You learn what the model says when prompted, not what a buyer sees while shopping.
- Trusting one run. Variation between runs is normal.
- Skipping the fact sheet, then arguing about what counts as wrong.
- Treating every omission as an emergency. Fix material errors first.
- Fixing a symptom in a chat window. You cannot edit a model's memory through its chat box; you change the sources it retrieves.
- Letting press coverage carry the correction. News is short-lived: in the half-life study, half of cited news pages were gone within 36 days, while landing and doc pages lasted far longer.
The short version
Write down what is true. Ask AI about you in ways a buyer would. Run everything more than once. Score the answers, trace each error to a page, and fix the page. Then check again next month. Half an hour is enough to find the worst problems, and most of them live on pages you already control.