EDITOR'S NOTE
Hey there 👋
Last week was about the sources AI reads when it describes your brand. This week is about what it actually says once it's read them.
Most AI visibility reports count mentions. That number on its own can mislead you.
In a study Omniscient Digital published on November 6, AI models named at least one brand in 97% of buying answers, but directly recommended one in fewer than half. You can show up constantly and still be the option the answer talks people out of.
Roche ran into a version of this. Its diagnostics business tracked 1,500 prompts across ChatGPT, AI Mode, Perplexity, and Claude with Semrush's enterprise AI tool. AI kept crediting Roche with innovation and global reach, then mentioned high costs and complexity in the same breath.
Roche isn't a financial brand, and its buyers are hospital labs, but the method transfers, and it's cheap to run at a smaller scale. This issue covers what Roche did, what the case study leaves out, and how I'd run the same four checks on your own brand, starting with 40 prompts instead of 1,500.
Let's go. 🚀
TL;DR 📝
Visibility and recommendation are different numbers. Omniscient found only 18.8% of brand mentions in AI answers were direct recommendations, and 72.9% of positive mentions didn't include a direct recommendation.
Roche tracked 1,500 prompts across four AI platforms and reports a 67.3% share of voice after three months, with 83% of sentiment positive and 14% neutral.
The case study gives no starting baseline, so it shows where Roche ended up, not how much it improved.
Roche's four steps were tracking the right prompts, reading how AI describes the brand, tracing the sources behind those descriptions, and benchmarking competitors.
I'd start with the same four checks, using 40 prompts a month, as long as they're the questions your buyers actually ask.
WORK WITH ME 🤝

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.
NEWS YOU CAN USE 📰

AI mentions your brand far more often than it recommends it. Omniscient Digital ran 300 non-branded buying prompts through ChatGPT, Claude, AI Mode, and AI Overviews once a day from September 30 to October 6, collecting 8,392 answers and 44,003 brand mentions. Two of the eight categories were checking accounts and credit cards. Models named a brand in 97% of answers but directly recommended one in 47.6%. Across all mentions, 18.8% were direct recommendations, 23.3% were conditional ("choose this if low fees are your priority"), and the rest were neutral options or warnings. ChatGPT recommended most directly (29.9% of mentions), while AI Mode did so least (10.7%). Among brands that appeared in both versions of a matched prompt, adding one buyer detail, like budget or team size, changed how 38% to 65% of them were recommended, for example, moving from listed to direct. Two caveats: an AI model did the labeling, and in a human check of 81 mentions it marked 16 as direct recommendations where the human marked 25, so read the percentages as approximate. It's still the clearest data I've seen on why mention counts aren't enough. [Source: Omniscient Digital]
Writesonic adds fact-checking and source tracing to its AI visibility tool. On November 5, Writesonic launched a Fact Check feature that flags wrong claims AI makes about a brand, like outdated pricing, and traces each one to the cited page. It also rebuilt its sentiment reporting to tie negative claims to specific passages, and added an Outreach Agent that drafts emails to the authors of pages you don't own, with every send needing your approval. The release also claims, from Writesonic's own study of 2,000 brands, that brands mentioned on YouTube or Reddit were more than three times as likely to appear in an AI answer. That's an association, and the release provides no methodology, so treat it with care. The useful part is that AI visibility tools are moving from counting mentions toward checking individual claims, which is the kind of review Roche did. [Source: PR Newswire]
HOW TO RUN ROCHE'S AI BRAND AUDIT ON A SMALLER BUDGET 🧠
Roche Diagnostics sells lab systems to hospitals, laboratories, and researchers, who ask detailed questions before they buy. Someone comparing diagnostic systems might ask an AI assistant which manufacturers offer the most automation or which systems suit a mid-sized lab. Roche can show up in both answers and be described very differently in each.
According to Semrush's case study, AI answers were already crediting Roche with global presence, innovation, and strong support, while also bringing up high costs and complexity. After three months of tracking and content work, Roche reported a 67.3% share of voice across its prompts and 97% positive or neutral sentiment.
Keep three things in mind before you copy it. It's Semrush's marketing for its own product. It doesn't say when the three months were, or how Semrush calculates share of voice. And it gives no starting figures, so there's no way to tell how much of that 67.3% came from the work and how much Roche already had.
The method is still useful. Here are the four steps, with how I'd run each one for a financial brand.
1. Track the Questions That Lead to a Purchase
Roche used Semrush's prompt generation to build 1,500 queries around its business priorities and positioning, then tracked them across ChatGPT, AI Mode, Perplexity, and Claude.
You can start much smaller. My suggestion is to start with 40 questions your buyers actually ask, split by stage: learning ("How do international business payments work?"), comparing ("Wise Business vs Revolut Business for a UK company"), and choosing ("Best business account for a freelancer who invoices in euros"). Omniscient's data shows why the split matters. Direct recommendations were 33.6% of mentions on recommendation-stage prompts but 11.6% on evaluation prompts.
A fintech can be visible on every "how does it work" question and missing from every "which one should I pick" question. If you only track one stage, you won't see it.
Run each question in a fresh chat with memory off, repeat it a few times, and log whether you were recommended, offered with a condition, listed as an option, or warned against. "Mentioned" isn't enough detail.
2. Read What AI Says About You, Word for Word
Once the tracking was running, Roche reviewed how the answers described the brand. Innovation, global presence, and diagnostics and automation leadership kept coming up as strengths. High costs and complexity kept coming up as drawbacks.
Do the same with your log. Quote the exact words. "Great app, but customer support is slow" and "great app" read very differently to a buyer, and an overall sentiment score can hide an important drawback.
Then separate tone from accuracy. An answer that calls your premium card expensive may simply be right. The question is whether the description is current, fair, and backed by a real source. According to the case study, Roche's response was targeted content to "balance the narrative" around costs and complexity.
Pay close attention to conditional recommendations. Omniscient found 23.3% of mentions came with a condition, like "choose this if low fees are your priority." Those conditions are AI's version of your positioning. If it keeps saying "good for beginners" and you're selling to active traders, that's the gap to work on.
3. Trace Every Recurring Claim to Its Source
Roche identified the third-party and Roche-owned pages AI cited most often, then looked for authoritative sources where it was underrepresented.
One line in the case study needs a closer look. It says "underutilized yet authoritative outlets on regulatory sites like the FDA were prioritized for content placements." It doesn't explain what "content placements" on a regulator's site meant in practice, so I wouldn't treat that part as a playbook.
For a financial brand, the legitimate version is checking that your regulator records are accurate: your FCA register entry, SEC or FINRA filings, or your state money transmitter licenses. Then, work through the sources you can actually influence, such as industry publications, comparison sites, review platforms, and your own help center.
For each recurring claim, check the pages cited next to it. If AI says your transfer fees are 1% and they're now 0.5%, look for an outdated page that still says 1% and fix it. If no cited page supports the claim, log its origin as unconfirmed, because AI can also infer a number or mix you up with another product.
4. Benchmark Competitors on What AI Says, Not How Often
Roche compared its AI visibility and positioning with competitors and found areas where they were ahead. Its response was more case studies, customer stories, and thought leadership to back up its strengths.
When you benchmark, go beyond counting mentions. For each competitor, note which attributes AI gives them, which sources it cites for those attributes, and which buyer conditions trigger a recommendation. Omniscient found that in project management software, Asana and monday.com both appeared in 87% of answers, but Asana was directly recommended in 38% of its appearances against 18% for monday.com.
If a competitor gets recommended for "small businesses" because it has five detailed SMB case studies and you have a single generic product page, that's a content gap with a clear fix. Publish the proof, with real customers and specific numbers, and make it easy to find.
What You Should Have After a Quarter
Run the same 40 questions every month, and you'll build up four things:
A recommendation rate, split by buying stage, that's more honest than a mention count
A list of how AI describes you, word for word, with the claims you need to fix flagged
A source for each recurring claim, separated into pages you control and pages you don't
A competitor table showing who gets recommended for which buyer situation, and why
THIS WEEK'S PROMPT 🧠

This takes the answers you've logged in step 1 and turns them into the analysis from steps 2 to 4.
The Scenario: You've run your 40 buyer questions from step 1 in fresh chats across ChatGPT, Perplexity, and AI Mode, and pasted each answer into a document with the question, platform, date, and any cited links. Now you need to know how you're being described and what to fix first.
The Prompt:
You're an AI search analyst reviewing how AI assistants describe my brand to potential customers.
Brand: [BRAND] Website: [URL] Category: [CATEGORY] Main competitors: [COMPETITOR 1], [COMPETITOR 2], [COMPETITOR 3] Target customer: [WHO YOU SELL TO] Verified facts: [FEES, MARKETS, REGULATOR, KEY FEATURES, ANYTHING THAT CHANGED RECENTLY]
Below are AI answers I collected. Each one is labeled with the platform, date, question, and any cited links. Work only from these answers and the verified facts.
For each answer, classify how my brand appears: directly recommended, recommended with a condition, listed as an option, warned against, or not mentioned. Quote the condition word for word if there is one.
Quote every statement about my brand and label it positive, neutral, negative, or mixed. Explain the label in one sentence.
Compare each factual claim with my verified facts and mark it Correct, Outdated, Wrong, or Unverifiable.
List the strengths and weaknesses that come up more than once. For each, say whether it looks like a fair description, an outdated one, or an unsupported one.
Do the same classification for each competitor, and list the attributes and conditions where a competitor gets recommended and I don't.
For each recurring claim, list the cited links that appear next to it. If you can browse, open each link and confirm whether it contains the claim. If you can't, say so.
Recommend five actions, ranked by how likely the issue is to put off a buyer.
Output a table with these columns: Platform | Question | How my brand appears | Quote | Sentiment | Accuracy | Cited source | Action.
Rules:
Don't add facts, sources, or competitor details that aren't in the answers or my verified facts.
Label any source you didn't confirm on the page as unconfirmed.
If the sample is too small to call something a pattern, say so.
AI ANSWERS: [PASTE HERE]
If 40 questions times three platforms is too much to paste in one go, run it in batches. Then, repeat the same question set every month and keep the tables.
My rule of thumb is to act on patterns that show up across three months and two platforms, with one exception: a single wrong claim about fees, eligibility, or regulatory status is worth fixing the day you find it.
WRAPPING UP 🌯
Track the questions that lead to a purchase, read what AI actually says about you, trace each claim to its source, and benchmark competitors on how they're described rather than how often.
Roche used 1,500 prompts and an enterprise tool. I'd start with 40 questions and a spreadsheet, and add more once you know which questions matter.
And when AI calls you expensive, check whether it's right before you try to change its mind.
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.



