Hey there 👋

Last week, I told you to run every audit prompt three times, in a fresh session with memory off. That's still right for catching errors. But it hides something that matters more for most financial brands, and that’s that your buyers aren't using fresh sessions with memory off.

They're asking from a specific city, in their own language, with months of chat history behind them. Some have connected their Gmail to Google's AI Mode, while others may have even connected their bank accounts to ChatGPT.

Each of those changes the answer. So the screenshot your team shares in Slack shows what one person saw once, and that's about all it tells you.

This week, I'll walk through how to measure AI visibility when there's no single answer to measure: repeat runs, markets, languages, regulated-category refusals, and the personal context you'll never see.

One number sets the scene. When the agency FTA Global asked ChatGPT and Gemini the same buying questions across 10 cities, 159 of the 206 brands they recommended showed up in only one country.

Let's go! 🚀

Alongside this newsletter, I help financial brands get found on Google, ChatGPT, Perplexity, and Google AI Mode, either directly or through my agency, Rise Up Media.

If you'd like help with yours, tell me what you're working on. It takes two minutes.

Work with me →

Where you ask from decides which sources AI cites. Mahe Chen and Nick Koudas ran about 31,200 queries through ChatGPT and Gemini, using 300 questions across 10 commercial topics, including finance and insurance. None of the questions named a place. ChatGPT adapted its sources to the location the app declared, down to the city, and ignored the network the request came from. Gemini followed the country of the internet connection, and it swapped out a noticeable share of its sources even for questions that had nothing to do with location. For "near me" questions, 96% of the businesses ChatGPT recommended appeared in only one of the 15 cities compared. It's a preprint posted October 16 and not peer-reviewed yet, but the practical point is clear: if your tracking tool doesn't control both kinds of location, its numbers describe wherever the tool happens to run. [Source: arXiv]

AI's top pick changes most weeks. The New York agency BusySeed asked ChatGPT, Claude and Gemini the same 160 buying questions every week from April 27 to August 3. The company named first changed from one week to the next 56.4% of the time. Gemini changed most often (61.2%), then ChatGPT (58.9%) and Claude (49.1%). Only 85 of 480 question-and-engine pairs kept the same leader for the whole period. It's vendor research run once a week, so it can't separate weekly drift from normal run-to-run randomness, but either way, one check a quarter isn't tracking. [Source: eCommerce Fastlane]

Gemini's Personal Intelligence is spreading to more places. Google is bringing Personal Intelligence to Gemini in Android Auto, where it can draw on Gmail, Calendar, YouTube, Maps, and search history. It's opt-in, and drivers choose which apps to connect. TechCrunch reports the feature has been available across the US since March and is now in beta in other markets. Every new surface is another place where the answer depends on who's asking. [Source: TechCrunch]

You can't make AI search give every buyer the same answer, and you can't see the answers it gives them. What you can do is measure how often you appear across the conditions that matter, and fix the ones where you don't.

There are six steps. The first sets up how you measure, the next three cover the conditions you can test, and the last two deal with the ones you can't.

In January, SparkToro's Rand Fishkin and Gumshoe's Patrick O'Donnell had hundreds of volunteers run 12 brand-recommendation prompts through ChatGPT, Claude, and Google's AI, 60 to 100 times each per platform. ChatGPT and Google's AI returned the same list of brands less than 1% of the time. The same list in the same order came up less than 0.1% of the time.

FTA Global found the same thing from another angle. Its team ran every question five times per platform, and the same brand came top in all five runs only 35% of the time.

So a single answer isn't a ranking. The useful number is share of answers: out of every run of a prompt, how many mention you, and how many put you first.

Run each prompt at least five times per engine per market, and log every run. "We appear in 4 of 5 ChatGPT answers in the UK and 1 of 5 in Germany" tells you where to work. "We're number two on ChatGPT" doesn't tell you anything.

FTA Global's Context Report, published September 17, ran 1,200 buying questions through ChatGPT and Gemini across 10 software categories, 10 cities and seven languages. Nearly three of the top five brands changed from one market to the next.

The money categories moved most. For payroll, accounting, payments, and HR software, an average of 4.4 of the top five brands changed between markets. For video conferencing and cloud storage, it was 1.5. In HR and recruiting, all five top brands changed compared with the New York answer.

The sources moved too. In Germany, 42% of the sources cited were local. In the UK, it was 28%, and in the UAE, 12%.

This is where Chen and Koudas's paper matters. ChatGPT takes its cue from the location the app declares, so a test run from your office in London tells you about London. Gemini takes its cue from the country of your internet connection, so you need a connection in each market, through a VPN or a tool that routes queries through local devices.

Before you trust any tracking dashboard, ask the vendor two questions: where do your queries come from, and do you set a declared location for ChatGPT? If they can't answer, the numbers describe their server.

FTA Global's Mumbai test ran in English, Hindi, and Hinglish, the mix of Hindi and English a lot of Indian buyers type in. The report doesn't publish a language-by-language breakdown, but if you sell in more than one language, assume the shortlist changes with it.

Don't translate your English prompts word for word. A German freelancer doesn't type "best business bank account for freelancers" in German. They ask about a "Geschäftskonto für Selbstständige." A Spanish buyer asking about sending money abroad might ask for the cheapest way to "enviar dinero" to a specific country.

Get a native speaker on your team, or a customer-facing colleague in that market, to write 10 prompts the way buyers ask them. Then check which local publications and comparison sites the engines cite. If 42% of German citations are local, a German placement does more for you there than another English article.

In June, Massive asked ChatGPT, Gemini and Copilot about three brands, including DraftKings, from the US, UK, Germany, Brazil and Japan. It's a tiny test, 270 queries in total, run by a company that sells the tool it used. But one result is worth knowing about.

Gemini refused both gambling prompts on every run from the UK and Germany, and answered them from the US, Brazil, and Japan. DraftKings' visibility in Gemini in those two countries was zero. ChatGPT and Copilot never refused.

Crossing a border barely changed what the assistants thought of a brand. It changed whether the brand appeared at all.

Massive didn't test financial products, so I can't tell you which ones get refused where.

My guess, and it's a guess, is that crypto, CFD trading, and high-interest lending prompts are the first ones to check. Log a refusal as its own result, separate from "not mentioned." A competitor can't beat you in an answer that doesn't exist, but you can't win it either, so the fix is in another channel.

Three products now feed private data into AI answers.

Google's Personal Intelligence launched in AI Mode on January 22 for AI Pro and Ultra subscribers in the US. It's opt-in, and it reads Gmail and Google Photos to tailor answers. In March, it opened to free users in the US, and it's now in beta elsewhere. Google warns it can make mistakes, such as connecting topics that aren't related.

ChatGPT's shopping research, launched last November, builds on what ChatGPT knows about you from past chats and memory. OpenAI's own example: if it knows you game, that shapes a laptop search.

And ChatGPT Finances, which matters most for this audience, links bank, card and brokerage accounts through Plaid. It launched as a US preview for Pro users in May and reached Free and Go users on October 2. OpenAI says it can save "financial memories" such as savings goals. It doesn't say whether it names specific providers, but a buyer asking about "moving idle cash into a higher-yield account" (OpenAI's example) now gets an answer that knows their balance.

Does that context change which products get recommended? A Cisco and Carnegie Mellon paper says it can. Across 325,000 experiments on 13 models, eight recommended pricier options when they believed the user was wealthy. In one test, an agent asked for the cheapest flight to Chicago picked a $91 Spirit ticket. After reading emails that mentioned a $680,000 401(k) and a Chase Private portfolio review, it picked a $601 United business-class seat while the $91 one was still available. The paper isn't peer-reviewed yet, but damn!

You can't track this directly. You can test it. Set up a few separate test accounts, give each one a buyer profile through custom instructions or a short chat history ("I'm a freelance designer in Berlin with €15,000 in savings"), and run your core prompts in each.

Then, make sure your pages answer for those situations in plain sentences. My view, and it's a view: when a model knows the buyer is a freelancer with irregular income, the page that says "for freelancers with irregular income, our account charges no monthly fee" has a better shot than the one that says "built for everyone."

Put it together in one grid. Rows are your core prompts. Columns are each engine in each market, plus your test profiles. Each cell is your share of answers out of five runs, with refusals marked separately.

Re-run it every month. BusySeed's data says the top pick changes more weeks than not, so a quarterly check misses most of what happens.

Track three numbers over time: your average share of answers, the markets where you're below 20%, and the local sources cited in each market where you're weak. The last one is your placement list.

When someone asks "where do we rank on ChatGPT?", the honest answer is a range: "We're in 60% to 80% of UK answers and under 20% in Germany." Report it that way and the conversation moves from screenshots to work.

This builds the market-by-market prompt bank from steps 2 and 3, using real search results instead of guesses.

The Scenario: You sell in several countries and need prompts that sound like local buyers, plus a list of the local sources worth checking in each market, before you run your first round of tests.

The Prompt:

Run it in a search-enabled chat, then run the prompts it gives you from a connection in each market, five times per engine.

These are the most popular AI tools we use at Rise Up Media. If you're not using them already, they're worth a look.

Full disclosure: some links above are affiliate links. If you sign up, we'll earn a small commission at no extra cost to you.

Run every prompt five times, test from each market and in each language, log refusals separately, build a few test profiles, and report the whole thing as a grid of ranges you refresh every month.

It's more work than a screenshot. It's also the only version that matches what your buyers see.

In FTA Global's test, 159 of 206 brands showed up in just one country. If you only test from your own, you don't know which of those you are.

Until next time, keep exploring the horizon. 🌅

Alex Lielacher

P.S. If you want your brand to show up in Google AI Mode, ChatGPT, and Perplexity, reach out to me.