Which questions about your brand will AI search answer without you? Finding and filling brand data voids

KEY TAKEAWAYS

1. A decline to answer is a void marker. Whoever publishes a specific answer first is likely to fill it.
2. Voids come in two kinds. Empty ones invite invention. Stale ones repeat your old prices and pages.
3. Placement beats volume. One post on a site an engine already cites outweighed 20 posts elsewhere.
4. Track cited and repeated as separate numbers, and expect deleted pages to linger in answers for days.

Most advice about wrong AI answers starts after the damage. Someone spots a bad price or an invented founder in ChatGPT, and the team works backwards to the source.

A better question comes first. Which questions about your brand does the web answer so thinly that an engine will take the first specific answer it finds? Those are your data voids. New research from October shows how quickly they get filled, and by whom.

What is a brand data void?

The term predates AI search. Michael Golebiewski coined it in May 2018 for search queries that return little or nothing, and he and danah boyd set out the risk in a 2019 Data & Society report. Their concern was that empty queries are easy to exploit.

For a brand, a void is any question a buyer might ask where few pages give a specific, current answer. Pricing tiers, refund terms, who founded the company, which integrations exist, how you compare to a named rival.

There are two kinds, and they fail differently:

• Empty voids. Nothing specific exists. The engine declines, borrows from a similar name, or uses the first page that offers detail.
• Stale voids. Old answers outnumber the current one. The engine repeats last year's price because most pages still say it.

The easiest way to spot an empty void is a refusal. Seer Interactive ran 1,562 prompts about its own brand on six engines daily from 1 May to 9 June 2026. The engines were right 95.9% of the time when they answered, but declined 31.2% of prompts. On comparison prompts they declined 80.3%. A decline is not a failure you can ignore. It is a slot nobody has filled yet.

How fast does a void get filled?

On a truly empty topic, within days. Researchers at Fudan University posted articles on the sites each AI search tool already cited most, then asked the same questions daily for a week. The tests ran in March and April 2026 and the paper appeared on 8 October.

Eight of ten tools cited an invented concept within seven days, including ChatGPT, Perplexity, Grok and an unnamed Google AI product. The team first mapped 17,211 citations and found that each tool leans on a short list of its own. The top 20 domains took between 20.5% and 70.8% of each tool's citations. On 15 of 22 publishing sites tied to those domains, the barriers to signing up and posting were low or medium.

Two details matter more than the headline:

• Placement beat volume. In paired tests on five holiday destinations, the post on a frequently cited site produced the first citation or planted detail every time. The best result from a rarely cited site needed 20 posts. Forty posts on a self-hosted WordPress site drew no citations in a week.
• Deletion did not end it. The team deleted all 27 articles after day 11. At least one tool kept citing the deleted content through day 15.

The study tested no real brands, and the samples for the English-language tools were small, 20 to 35 answers each. Treat the rates as a direction, not a forecast.

Are AI answers fragile or stable?

Both, depending on the question. BrightEdge tracked thousands of prompts in nine industries across five engines and found 96.8% of cited domains had no change from one week to the next in February 2026. That looks like the opposite of the Fudan result.

The two studies measured different ground. BrightEdge watched established industry questions, most of which already have many pages answering them. Fudan built empty topics on purpose. Its own data shows the gap. Once a tool had started citing the planted posts, answers on the invented concept cited them 62.9% of the time. On an opinion question that already had plenty of discussion, the rate fell to 18.8%.

So the evidence supports a narrow claim. Answers are stable where coverage is dense and easy to move where it is thin. Your category head terms are probably dense. Questions about your own company often are not.

Is the bigger risk invented facts or old ones?

For most established brands, old ones. Timmermans Media asked 185 factual questions about 37 Dutch brands on ChatGPT, Gemini and Perplexity in July 2026, and counted an answer wrong only if two of three runs were wrong. Of the errors, 62% were outdated and 11% fabricated. Time-sensitive facts such as prices and terms were wrong 6.8% of the time, against 2.5% for stable facts.

Profound's August 2026 analysis of more than 158,000 brand claims points the same way. Pricing and billing made up 12% of the claims checked but 24% of the inaccurate ones. And 54% of brands had at least one wrong claim that cited their own site. Your old pages are part of the stale void.

Small brands face more invention. Searchable's study of 165 London businesses found a fabrication rate of 5% for small firms, against 2% for large ones. Timmermans found no small-brand penalty, but its sample held only seven small brands and its majority rule filters out one-off inventions. The two results fit together. Where almost nothing exists, engines invent. Where something old exists, they repeat it.

Does being cited mean being believed?

No. Cited and repeated are separate outcomes, and you should count them separately.

In the Fudan test, Perplexity cited the planted posts in 13 of 35 answers yet repeated none of their planted details. Grok cited them in 30 of 32 answers and repeated details in 53%. ChatGPT cited them in 7 of 20 and repeated details in 10%.

Ahrefs reached a harsher verdict on Perplexity in a December 2025 test with an invented brand. After three planted stories, Perplexity and Gemini repeated misinformation in roughly 37% to 39% of answers, and ChatGPT in under 7%. Roger Montti later counted 49 of the 56 questions as leading, since they assumed false facts. That explains much of the gap. Leading questions pull repetition, and buyers do ask them.

One finding from Ahrefs holds either way. Its author concluded that "the most detailed story wins, even if it's false." The brand's own FAQ declined to give details where the fake sources offered names and numbers. Vague official answers lose to specific unofficial ones.

What does a void audit look like in practice?

A hypothetical example, with invented figures. A 30-person scheduling software company writes 20 questions about itself and runs each on four engines, for 80 answers. It scores every question on two counts: how many independent pages answer it specifically, and what the engines did.

Four of five questions are voids. The founder and comparison rows carry the most risk, because the first specific answer published anywhere will likely become the answer. The pricing row is the stale case, and the fix starts on the company's own site.

Which void needs which fix?

How do you find and fill your voids?

1. List 20 to 40 questions a buyer might ask about you, not about your category. Include pricing, terms, people, integrations and named comparisons.
2. Run each on every engine you care about, several times. Log declines, wrong answers and right answers. Sampling guidance is here.
3. Count specific pages per question. Search for each answer yourself. Zero or one current page means a void.
4. Mark each void as empty or stale, and rank by stakes.
5. Publish specific answers on your own site first. Use numbers, names and dates, not an invitation to get in touch.
6. Place the same facts on a site each engine already cites. Check the cited domains in your own answers. One placement there can do more than many elsewhere. That is what the placement marketplace is for.
7. Recheck weekly at first, and log cited and repeated separately. A citation that does not carry your facts is not yet a fix, much like a citation that never names you.

What should you not do?

Do not seed fake independent posts, even to fill a void with true facts. In May 2026 Google added attempts to manipulate generative AI responses in Google Search to its definition of spam. Its AI optimization guide also says that seeking inauthentic mentions helps less than it seems.

And do not expect feedback buttons to fix things. Google lets you report an AI Overview through the thumbs-down menu, but its help page describes no outcome or timeline. OpenAI's publisher FAQ gives no timing for changes. In a 2024 complaint filed by noyb, OpenAI said it could not correct false data in the model and could only block it. Even the source fixes take time. Google says a recrawl can take a few days to several weeks. Business Profile edits usually clear in 10 minutes but can take up to 30 days.

In digital authority terms, filling voids touches four of the five pillars. Your own pages and profiles build brand entity and search presence. Placements on cited sites add press coverage. Fewer declines and fewer wrong answers show up as AI visibility. Start with the questions engines refuse to answer today. Those are the ones someone else can answer for you tomorrow.

Questions

What is a brand data void in AI search?

It is a question about your company, such as pricing, founders, policies or how you compare, that few pages answer specifically, or that old pages answer wrongly. AI engines either decline to answer or fill the gap from whatever they retrieve, including forum posts, stale reviews or your own outdated pages.

How quickly can a new post change what AI search says?

On empty topics, within days. Fudan University researchers posted articles on sites the tools already cited and found 8 of 10 AI search tools cited an invented concept within seven days in March and April 2026. On a topic with existing discussion, citation rates were far lower.

Are AI engines more likely to invent brand facts or repeat old ones?

Mostly repeat old ones. A July 2026 Dutch study found 62% of brand errors were outdated and 11% fabricated, and Profound found pricing made up 24% of inaccurate brand claims. Invention is more likely where almost nothing exists, which is where small brands sit.

Can I ask ChatGPT or Google to correct a wrong answer about my brand?

You can send feedback, but no vendor commits to fixing a specific answer or gives a timeline. Google's AI Overviews take reports through the thumbs-down button. In a 2024 data protection complaint, OpenAI acknowledged it could not correct false data in the model itself. Fixing the sources the engines retrieve is the dependable route.

Sources

1. arXiv via alphaXiv, Liu et al., From public posts to AI-search citations. Measuring the fragility of AI search alphaxiv.org
2. Search Engine Journal, AI search tools cited a made-up concept within a week, paper finds searchenginejournal.com
3. Relevant Audience, AI search tools cited a made-up concept within a week, Fudan University paper finds relevantaudience.com
4. Microsoft Research, Data voids. Where missing data can easily be exploited microsoft.com
5. Seer Interactive, AI brand accuracy study seerinteractive.com
6. Timmermans Media, AI hallucination rates by model for Dutch brands timmermansmedia.com
7. Profound, Where do inaccurate AI claims come from? tryprofound.com
8. Ahrefs, I ran an AI misinformation experiment ahrefs.com
9. PPC Land, What Ahrefs' fake brand experiment actually proved about AI search ppc.land
10. BrightEdge, AI search citations. How much do they really change week to week? brightedge.com
11. IT Brief UK, London SMEs suffer AI misinformation in search tests itbrief.co.uk
12. Google Search Central, Spam policies for Google web search developers.google.com
13. Google Search Central, Optimizing for generative AI features in Google Search developers.google.com
14. Google Search Help, AI Overviews support.google.com
15. Google Search Central, Ask Google to recrawl your URLs developers.google.com
16. Google Business Profile Help, Understand what happens to your Business Profile edits support.google.com
17. OpenAI Help Center, Publishers and developers FAQ help-lb.openai.com
18. noyb, ChatGPT provides false information about people, and OpenAI can't correct it noyb.eu

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →