Asking ChatGPT About Your Brand Is the New Googling Your Own Ads

Googling your own ad is a lousy way to measure PPC. Asking ChatGPT about your own brand is the same mistake with a newer interface.

I saw the ad version constantly in my PPC years: someone searched for their company from the office, saw the ad, and relaxed. That check said nothing about auction share, search terms, or cost per acquisition. It said Google could show that person their own ad in that moment.

Now a founder logs into ChatGPT, asks for the best company in their category with their town attached, sees their name, and decides they are fine across ChatGPT, Gemini, and Google AI Overviews. I get the appeal. The answer is immediate and looks like the customer experience. But one personalized check measures the check, not the market. The mistakes that follow are familiar to anyone who has watched a screenshot become a strategy.

“I asked ChatGPT and it mentioned us, so we’re fine”

This is the most common belief I see. You ask from your own account, on your own laptop, and ChatGPT knows your company. It feels like a customer asking. It is not.

When Memory is enabled, ChatGPT can use preferences from past chats, files, custom instructions, and connected apps to personalize responses. Your earlier conversations and corrections can tilt the answer toward you. That is useful when you want help with your work. It is a terrible control when you want to know what a buyer sees.

The cost is false reassurance. You stop investigating because a session that already knows your name supplied it back to you. My fix is the same one I used for ad previews: stop treating your own account as a neutral observer. Test logged out in an incognito window, and check with web search on and off separately. Those checks answer different questions about what appears without live retrieval and what appears when the system can search. If you show up in only one state, record that condition. Do not turn it into a claim about visibility everywhere.

Takeaway: never trust a mention you only saw while logged in.

“We should track prompts with our brand name in them”

This one feels like diligence. You build a list of prompts like what does [your brand] do and is [your brand] any good, run them weekly, and watch your name come back.

Of course it comes back. You put it in the question.

Branded prompts check representation accuracy, not discovery. They can tell you whether an answer describes your business sensibly. They cannot tell you whether a buyer who has never heard of you will find you. For many local and B2B businesses, that unbranded moment is the point: someone describes a problem and lets the model make a shortlist.

The cost is a visibility score that flatters you while missing new demand. A score of 100% on prompts containing your name and near zero on prompts without it is not demand coverage. It is a mirror. Keep branded prompts if you need to inspect how you are described, but give unbranded buyer prompts more weight when the question is whether new customers can discover you.

Takeaway: do not count being named in your own question as evidence that you made the shortlist.

“A mention is a citation”

I used to count any PPC sighting as a win, too. Impression, click, lead: it is easy to blur them until you tie spend to closed revenue. AI answers need the same discipline because a name in the text and a link to your site do different jobs.

• A mention names your brand in the answer.
• A citation links to your domain as a source behind a claim.
• A recommendation endorses you as a choice.

Do not collapse those three events into one visibility number. An answer can use facts from one page while naming another brand. A Semrush finding cited by ZenMedia reports that 61.7% of AI citations never name the linked brand. You can supply the source without getting named, or get named without giving the reader a link. A recommendation tells you something else again: whether the answer actually favors you.

That distinction matters when you decide what to fix. If you are named but not linked, another mention will not necessarily solve the missing path to your site. If you are cited but a competitor gets the recommendation, a citation count alone will hide the loss. Gemini and ChatGPT can behave differently on mentions and citations, so keep the columns separate by engine as well.

Takeaway: log mentions, linked citations, and recommendations separately. A name in an answer is not automatically a source, an endorsement, or a visit.

“If it showed up Tuesday, it shows up”

One clean screenshot feels like proof. I distrust any measurement that only works once.

AI answers re-compose when you ask again. Wording shifts, sources rotate, and a competitor can slide in because you added one word like cheap or near me. One analysis puts single-answer reliability near 0.01; brand identity explains only 1.5% of the variance, while re-running the same prompt explains 34.8% and query language another 26.5%. However tidy the screenshot looks, a single run leaves a lot of room for noise.

The tempting response is to hammer the same question until the answer settles down. That has limits, too. In the cited analysis, reliability improves by only 0.0003 past the fifth repeat. Breadth across prompts and engines raises confidence more, lifting reliability toward 0.36. You need repeats, but you also need different ways a buyer might ask.

The cost of skipping either is chasing a Tuesday appearance that vanishes on Wednesday, or treating a missing name on one run as a crisis. Five runs across ten buyer prompts on three engines is a start of data, not a guarantee that every future answer will match.

Takeaway: repeat the check and widen the prompt set. If you cannot say how often you appeared, you have not measured visibility.

“Google rank tracking but for AI”

In paid search, I lived inside positions and impression share. A position number pointed back to bids, Quality Score, and auctions you could inspect. An AI rank tracker showing position 3 and a little green arrow borrows that familiar language. It borrows more confidence than the measurement deserves.

Answer engines compose fresh responses rather than preserving a stable ranked list with a dependable position 3. What you can inspect is whether you are named, cited, recommended, or prominent in a particular answer. A top Google rank does not guarantee inclusion; these engines can draw from pages outside the top 10 or pages that did not rank there at all.

The cost is wasted work. A business can spend its effort moving a page from rank 6 to rank 4 while the buyer prompt it cares about still produces no citation to that page. The rank improved. The answer did not. Do not use a familiar position number to hide an unfamiliar problem.

For each buyer prompt, I want the plain, slightly uglier record: did we appear, did we get the link, and were we the pick or paragraph seven? That gives you something to investigate instead of an arrow to admire.

Takeaway: stop asking where you rank. Ask how often you are named, cited, and recommended.

“Gemini and AI Overviews are the same engine”

They both have Google attached, so treating them as one system sounds efficient. It can also leave a large gap in your report.

An analysis of 383,188 citations across 64 ecommerce brands found median domain overlap of just 15.8% between Gemini and AI Overviews. Only 2 of those 64 brands shared the same number one source in both. In that analysis, AI Overviews drew more heavily on social content, at 19.83% versus 2.63% for Gemini, while Gemini favored editorial pages and year-in-title pages at more than three times the rate.

Same company. Different source patterns. If you measure one surface and assume the other follows, you can miss the very absence you are trying to find. A plumber who looks solid in AI Overviews through Reddit threads and YouTube clips may not appear in Gemini for the same kind of buyer question, where a structured how-to or comparison page may matter more.

The cost is an apparently healthy blended score that gives you no clue where the weakness sits. Track Gemini and AI Overviews as separate lines for the same prompts. Yes, that doubles those rows in the sheet. It also lets you see which surface needs attention instead of averaging the gap away.

Takeaway: if your report folds both into one Google AI number, ask to see them split.

“I can check a few prompts whenever I have time”

An informal check is easy to start and hard to compare with anything later. Change the prompt, switch engines, forget whether you were logged in, and this week’s better-looking answer might reflect your method rather than your visibility. That is how a screenshot acquires a trend line it never earned.

If I had to tape a measurement checklist to a monitor, it would look like this:

1. Choose 30 to 50 buyer prompts written the way customers talk. Make most of them unbranded; keep branded prompts for checking how the business is represented.
2. Run the prompts on separate engines and surfaces: ChatGPT, Gemini, and AI Overviews. Record the conditions of the check rather than mixing logged-in and logged-out results.
3. Repeat each prompt. The Darwin audit approach starts with 30 to 50 buyer prompts run 5 to 10 times per platform. Repeats expose how much a single answer can wobble.
4. Score each answer three ways: named, cited with a link, and recommended. Keep the prompt and engine beside the score so you can find the pattern again.
5. Compare on a schedule. Brand Visibility Score is answers mentioning the brand divided by total answers, times 100. Track that alongside citation rate, share of voice against named competitors, and sentiment on your top-intent queries weekly or monthly.

That is sheet work. It is also what lets you tell a content problem from a sampling wobble before you spend money trying to fix it. I would rather hand a beginner a fixed protocol like the 30-prompt test for AI search visibility than tell them to freestyle a few questions whenever they remember. Same prompts, same surfaces, same scoring, repeated on schedule.

With groas, the work can continue past that baseline: content, technical fixes, and citation work can be logged against the same measurement instead of ending in a PDF that says to “improve authority” and leaves you to guess how. A number you can defend beats a screenshot you can frame.

Takeaway: if your report cannot show its prompt list, repeat count, and split between mentions and linked citations, it is not tracking. It is anecdotes with a logo on top.

“If sentiment drops, we should fix it immediately”

This is the belief I am still arguing with myself about. I track sentiment because a citation that warns people away from you can be worse than no citation. An answer can link to your pricing page as the source for a claim and still tell the reader to pick a competitor for reliability or support. That looks like a citation win in a dashboard and a lost job in real life.

I have seen answers name a brand repeatedly while the verdict goes to someone else. Counting names and links alone misses the part that decides money. The practical line I use is: a mention without a recommendation is inventory, not income.

What I have not settled is how often a small business should act on sentiment movement. With 30 prompts run five times each, you get 150 answers per engine. Two or three negative sentences can shift a monthly sentiment share by points that look important and mean nothing. That is sampling noise dressed as a trend.

For now, I log sentiment weekly but act monthly, and only when the same complaint repeats across different prompts and engines. Say you spend $20k a month on search and AI answers keep flagging your install times: that pattern is worth a page fix. One grumpy paragraph on one Tuesday is not.

Same lesson as the ad preview, all over again. Stop asking the machine if it likes you. Build the control, run it on schedule, and fix what repeats.