What changed · Citation sources
You are not in the answer because you are not in the sources
Brands rewrite the page they control and wonder why nothing shifts. The answer was assembled somewhere else, out of sources they have no presence in.
By Priyadharshini Krishnan, Co-Founder & Chief AI Officer · Updated 16 September 2026 · 7 minute read
When an assistant answers a category question it draws on a small set of domains it treats as reliable for that category, and your own website is rarely the most influential one. Semrush analyzed over 100 million citations across ChatGPT, Google AI Mode and Perplexity over thirteen weeks and found Reddit, Wikipedia, LinkedIn, Forbes and Medium leading, with the mix differing sharply between engines. So a good deal of the work sits off your own site: being present, accurately described and genuinely useful where your category gets discussed. Forums, review hubs, trade press, directories, reference pages. It is slower than editing your own content, which is exactly why most brands never get to it.
How narrow is the source set?
Narrow, and unstable. Semrush sampled 230,000 prompts weekly between July and October 2025 and analyzed more than 100 million citations. Reddit, Wikipedia, LinkedIn, Forbes and Medium led overall.
The more useful finding is how differently the engines behave. Google AI Mode leaned toward properties Google owns or partners with, LinkedIn and YouTube among them, and cited Wikipedia around 2% of the time. Perplexity cited Reddit heavily and stayed comparatively steady. ChatGPT did neither: Reddit fell from roughly 60% of its citations in early August to roughly 10% by mid September, Wikipedia from 55% to under 20% across the same weeks. No publisher did anything to bring that about.
What this means in practice
Do not build a strategy on one platform. A brand that poured everything into Reddit presence through August was materially less visible in ChatGPT by September, having done nothing wrong. Spread across several source types and assume the weighting between them will move without warning or explanation.
Which sources are worth working on?
Start from your own category, because the general list is where all the volatility lives. Four types are worth auditing.
- Reference and directory pages for your category. Unglamorous, comparatively stable, and often where a model picks up the basic facts about who exists at all. Being missing or wrongly described here is a quiet handicap that never announces itself.
- Review and comparison hubs. Where buyers take comparison questions, and where models go to answer them. Presence matters more than score, because an absent brand cannot be recommended at any rating.
- Trade and industry press. Slower and harder, and the most durable of the four, since it is the source type least likely to get deweighted as unreliable.
- Community forums where your category actually gets discussed. High influence, and the place brands most often behave badly. A promotional post does not help and frequently hurts.
How do you find where your category is discussed?
Work backwards from the answers. Ask the assistants your category questions and write down what they cite. That list is the map, specific to your category, and far more useful than any published study including the one cited above.
Then set it against where competitors appear and you do not. The gap is your work list in priority order, and it is usually shorter than people expect. Most brands find three or four places carrying most of the difference.
What does good work on a source look like?
Much the same as it looked before any of this existed, which is the unsatisfying part of the answer. Being useful where people are asking questions, under your own name, without pretending to be a happy customer.
The failure modes are worth stating plainly because both are being sold right now. Astroturfing gets caught, increasingly by Signals themselves. Bulk placement on sites that exist to sell links gets those sites deweighted, and takes your mentions down with them. Each converts a budget into a liability.
The version that lasts is slower. Correct and complete entries wherever your category is cataloged. Real presence in the two or three communities that matter, built over months. Working relationships with the publications that already cover your industry. None of it is new. What is new is that it now feeds an answer as well as a reputation.
You control your website completely and your citations barely at all. The engines weight them the other way round.
What to do this week
- Run your ten category questions and list every domain cited. That list is your real competitive set, and it will contain surprises.
- Check whether you are present, and correctly described, on the reference and directory pages in it.
- Mark the sources citing competitors but not you, and pick the two most winnable this quarter.
- Check that nothing you are currently doing depends on one platform staying weighted the way it is today.
If you are an agency
Here an agency retainer is worth more than software, because this is relationship and placement work that a tool can point at but cannot do. The deliverable worth selling is the gap list: which sources answer this client's category questions, which of them cite competitors, and which are realistically winnable inside a quarter. That is a scope with a visible before and after, and it does not require promising a ranking that nobody controls.
How we do this ourselves
Our AI search visibility module records the sources cited in every answer it collects, so the domain list above gets built as a by-product, and it tracks which competitors turn up in which sources. Where a gap is worth pursuing it will draft the outreach, and it verifies what happened three ways: whether the piece was published, whether it was then cited, and whether the brand now appears in answers to the questions that gap was about.