Ask ChatGPT, Claude, Gemini or Perplexity for "the best project management tool for a small agency" and you get a short list of names, usually with a sentence on each. Nobody typed a keyword into a search box and nobody clicked ten blue links. The assistant made a call on which brands deserved to be on that list.
That call is not random, and it is not a black box you can do nothing about. It comes from a handful of mechanisms that work together. If you understand them, you can work out why your brand shows up in some answers and not others, and what to change. This guide walks through each mechanism, what is publicly known about it, and where the honest limits of our knowledge are.
Two sources of knowledge: memory and retrieval
Every assistant answer draws on one or both of two things.
Training data (memory). A language model is trained on a large snapshot of text. During training it learns which names go with which categories, which products are described as alternatives to each other, and how people talk about them. When you ask a question without any live lookup, the model answers from this learned memory. Brands that appeared often, in clear and consistent contexts, in the training corpus are easier for the model to recall.
Memory has two properties that matter for marketers:
- It lags. A model reflects the world as of its training cutoff, plus whatever later fine-tuning added. A product you launched last quarter may simply not exist for a model answering from memory.
- It is slow to change. You cannot edit it directly. You can only change what the world says about you and wait for future models to learn it.
Live retrieval (search grounding). Most major assistants can also search the web, pull in pages, and write an answer grounded in what they found. Perplexity is built around this. ChatGPT, Claude and Gemini use it when a question looks like it needs current information, and Google AI Overviews are generated alongside Google's own search results. When retrieval is active, the answer is shaped by whichever pages the system fetched and judged relevant.
Retrieval is where most of the short-term opportunity lives. A page you publish today can be fetched tomorrow. That is why the question "can the assistant's crawler or search system reach and understand my page?" matters so much, and we cover it in our guide to AI crawlers.
Which mode is the assistant in?
You can usually tell by the response. Answers with numbered source links or footnotes were grounded in retrieval. Answers with no sources, especially for evergreen questions, probably came from memory. The same prompt can flip between modes depending on the product, the user's settings, and how the question is phrased. Questions with words like "latest," "2026," "pricing" or "reviews" tend to trigger a search.
What shapes a recommendation
Whether the answer comes from memory or retrieval, a few factors keep showing up.
1. Entity clarity
An assistant has to know what your brand is before it can recommend it. "Entity clarity" means the web consistently says the same thing about you: your name, what category you are in, who you serve, and what you are not.
Things that blur an entity:
- A brand name that is also a common word or another company's name.
- Different descriptions on your homepage, your LinkedIn page, your directory listings and your press mentions.
- Products with unclear boundaries, so the model cannot tell whether you are a "CRM," a "sales engagement platform" or a "marketing suite."
Things that sharpen it: a plain-language one-sentence description used everywhere, a clear About page, structured data on your site, and consistent naming across profiles and listings.
2. Consensus across sources
Assistants are cautious about claims that only one source makes. When several independent places agree that a brand is a good fit for a use case, such as review sites, editorial roundups, community threads and the brand's own documentation, the model has more reason to repeat it. When sources conflict, the answer tends to hedge or leave the brand out.
This is the biggest practical difference from classic SEO. You can rank first for a keyword with a strong page of your own. But if nobody else mentions you, an assistant summarizing "what do people recommend?" has little to summarize.
3. Fit between the question and the page
In retrieval mode, the system is matching a prompt to passages. Pages that state the answer directly, in the same terms the user used, are easier to pull in and quote. A page titled "Best invoicing software for freelancers" with a clear comparison is a better match for that prompt than a generic feature page. We go deeper on this in how to structure content that gets cited.
4. Source trust and type
Not all sources carry equal weight. In our experience tracking answers, assistants lean on a recognizable mix: publisher and editorial sites, official documentation, review and comparison sites, and user-generated content such as forum threads and video. Which types dominate varies by industry and by engine, which is exactly why you should measure it for your own prompts instead of assuming. We cover the community side in Reddit and YouTube in AI answers.
5. Freshness
For retrieval, newer pages can win on questions where recency matters. Out-of-date comparison pages and old pricing are a common reason a brand is described wrongly or skipped.
Why the same question gives different answers
Language models generate text by sampling. Ask the same question twice and you can get a different list, a different order, or a brand that appears in one and not the other. Add differences between engines, between countries and languages, and between logged-in and logged-out users, and a single check tells you very little.
That has a direct consequence: treat any one answer as a sample, not a verdict. We explain how to measure around this in measuring AI visibility.
What you can influence, and what you cannot
| Lever | Your control | Time to effect |
|---|---|---|
| Pages on your own site that answer target prompts | High | Days to weeks, once crawled |
| Letting AI search crawlers reach your content | High | Days |
| Consistent entity description across the web | Medium | Weeks to months |
| Third-party reviews, roundups and community mentions | Low to medium, earned not bought | Months |
| What is baked into a model's training data | Very low | Next model release |
Be wary of anyone who claims they can guarantee a spot in an AI answer. Nobody outside the model providers controls the final selection. What you can do is remove the reasons you would be skipped.
A practical diagnostic
Run this on a short list of prompts that matter to your business:
- Write 10 to 20 prompts a real buyer would type, including comparisons and "best for" questions.
- Run each one several times on each engine you care about.
- Record whether you are named, in what position, and which URLs are cited.
- For prompts where a competitor appears and you do not, open the cited sources. Look for the pattern: a listicle you are missing from, a Reddit thread, a page that answers the question more directly than yours.
- Fix the specific gap, then re-measure.
The citations are the most useful part. They tell you what the assistant actually read, which is the nearest thing to a view into why it chose the brands it chose.
How Citeflare helps
Citeflare runs your tracked prompts on ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews on a schedule, records which brands are mentioned and where, and lists the sources each answer cited. You can see the gap between you and your competitors prompt by prompt, instead of guessing from a single answer. Start tracking your prompts or create a free account to see where you stand.