Saga Torseke speaks for Profound, whose FactCheck product traces inaccurate brand claims in AI answers back to their sources. In this email interview with MarTech Edition, Torseke explains where those errors come from, what a brand can change once the source is found, and who inside a marketing team should own answer-engine accuracy.
Your research found 47% of LLM response content is commentary the user didn’t ask for. Where does most of the inaccurate brand information come from?
Much of it comes from that unsolicited commentary rather than the direct answer. When someone asks, “What’s the best card for travel?” the model answers, then keeps going. It explains why it picked something, builds a comparison table, adds fee and terms caveats, and ranks options. Those additions carry some of the most specific claims about a brand, and specifics are where errors often occur.
Our data shows how consistent this behavior is. The most common type of unsolicited content across all three major models is rationale for the answer, such as “extensive e-commerce features and integrations.” It makes up 9.7% of ChatGPT’s editorial content, 8.8% of Claude’s, and 8.6% of Gemini’s. These are claims about your product that you didn’t write and may not agree with.
The highest-risk categories are the ones full of hard facts. Comparison tables account for 8.0% of ChatGPT’s unsolicited content and 8.3% of Gemini’s. Fee and terms caveats, like “Annual Fee: $95,” account for another 5.3% on ChatGPT and 4.8% on Gemini. Rankings show up in 4.5% of ChatGPT’s unsolicited content and 4.7% of Claude’s. A table that lists your pricing next to a competitor’s, using information from a two-year-old affiliate article, is how a brand ends up misrepresented in an answer it technically “won.”
The models also differ. Claude builds comparison tables far less often (3.0%) but leans more on follow-up referrals (5.5%) and context-dependent advice (4.7%). Gemini ranks options less often than the others (2.7%). So the risk to your brand looks different on each engine.
Inaccurate claims can originate from third-party pages, including review sites, affiliate roundups, and outdated comparison posts. These sources can continue influencing AI answers even after a brand has updated its own website.
WHOOP found accuracy issues in 7.9% of AI responses that cited the brand. What rate do you see across customers, and what pushes it down fastest?
The rate we see across customers varies considerably. WHOOP’s rate of 7.9% offers a useful example of the accuracy issues brands encounter, although the exact rate depends on the brand, industry, and questions being asked.
What pushes it down fastest is often fixing the handful of sources that drive most of the errors. Errors aren’t necessarily evenly distributed. A small number of pages, sometimes one outdated comparison article or a stale pricing page, can account for a large share of the inaccuracies. Correct or replace those, make sure your own pages state the facts clearly and consistently, and you can often improve accuracy faster than you would by simply publishing more content.
Once FactCheck traces an error to its source, what can a brand actually change?
More than most people assume. It depends on who owns the source:
- Your own pages: update pricing, plans, features and specs, remove contradictions between pages, and state key facts plainly so a model can lift them accurately.
- Third-party pages: ask publishers, review sites and affiliates to correct outdated claims. Many will, especially when you give them current, verifiable information.
- Gaps: if a model is filling in a blank with a guess, publish the authoritative answer so it has something better to cite.
Then you monitor, because models refresh on different timelines and you want to confirm the correction actually shows up in answers.
- Your own pages: Update the page. AEO Guide ch. 5 calls this the on-page fix. Keep pricing and other fast-changing details consistent everywhere.
- Third-party and earned pages: Contact the publisher. An outreach agent can draft the note, and the Noble integration can contact the publisher for you. The guide ranks this off-page work first, because you can’t out-publish a wrong answer the engines keep finding elsewhere.
- Competitor pages: Create your own content that states the correct fact. FactCheck can launch an Agent for this.
- Your ground truth: Update the Knowledge Base, for example before a launch (FactCheck FAQ). Report claims wrongly marked inaccurate.
FactCheck keeps rechecking after a fix, so you can watch the score move. Freshness helps: half of the content engines cite is under 13 weeks old.
Who inside a marketing team should own answer-engine accuracy?
Product marketing, if it exists at a company, is often the natural owner. PMM already owns the source of truth for what the product does, how it’s priced, how it compares, and how it should be described. Answer-engine accuracy extends that responsibility to a new channel.
Execution is shared. Content and web teams update owned pages, PR and communications handle third-party corrections, and in regulated categories, legal and compliance teams need to be involved. Marketing leadership should establish accountability across those functions, with PMM responsible for making sure AI describes the product correctly.
What do you expect marketers at the summit to ask about answer engine optimisation, and what will you tell them?
- “Isn’t this just SEO?”
- There’s overlap, but the unit of competition is different. You’re competing to be included and described accurately in a single synthesized answer, rather than simply earning a click from a ranked list. Visibility, citations, sentiment, and accuracy all matter. Traditional SEO has generally focused on rankings, traffic, and conversions, while AEO also requires marketers to measure what AI actually says about their brands.
- Visibility is the dimension most similar to SEO, but it’s only one part of AI search. That’s why we’re also looking at sentiment and accuracy.
- “Which model should I focus on?”
- All of them, because they behave differently. Our data shows Claude builds comparison tables far less often than ChatGPT or Gemini (3.0% versus about 8%). Gemini ranks options less often than either. Your exposure and your errors will look different on each engine.
- “Can I actually influence what AI says about me?”
- Yes. You can’t edit a model directly, but you can influence the sources it draws from. Brands can measure what answer engines say, trace errors to their sources, and correct inaccurate information. Publishing more content alone won’t necessarily fix an error that originates elsewhere.
- How do I prove ROI?”
- Attribution in any kind of marketing is difficult, especially when there is no click, which is often the case in AI search. But customers have reported increases in referral leads and revenue from AI channels after working with Profound, as well as significant time savings from using Profound’s agents.
- WHOOP, for example, reported a 6.6% increase in visibility over six months, 56% presence on top prompts, and more than 100 hours saved. Lake.com also reported a 5× increase in branded search and direct traffic after investing in AEO. Looking at those outcomes alongside changes in AI visibility and accuracy can help marketers understand the impact of their efforts.
- “Who owns it?”
- AEO should have ownership at the marketing leadership level rather than being buried within SEO. It requires a top-down mandate and cross-team OKRs, with SEO helping coordinate execution. Under that model, product marketing manages landing pages and product content, while PR, partnerships, and communications handle third-party outreach. Product, Customer Success, Brand, and Demand Generation also contribute. The exact ownership structure depends on the company’s organization, business model, and goals.