Answer engine optimization, for communications teams
AEO is the work of understanding what AI agents say about you when someone asks, and what moves it. It belongs on a communications desk because the answer an agent gives is assembled from the same material that shapes public opinion: press coverage, review sites, community forums, reference databases, and your own site.
AEO, GEO, AI visibility
The terms get used interchangeably, and mostly they are.
Search engine optimization competes for a position in a list of links. Answer engine optimization and generative engine optimization both concern the answer itself: which brands an agent names, which sources it cites, and how it describes you. We use AEO and GEO to mean the same thing, and so does most of the market.
Nothing below depends on the distinction.
Why this lands on a communications desk
An agent answering a buyer's question is reading press coverage, trade titles, review sites, forums and reference databases, then assembling a view. That is a media problem before it is a technical one.
AI is an audience, not a marketing lever. Arguably it is now the most important one, because it reaches your buyer before your buyer reaches you.
In one read of a large financial firm, the headline figure the agents quoted for the firm's size differed by about a third from one answer to the next, with as-of dates that contradicted each other, depending on which source each answer leaned on.
Read that way, what an agent says is a proxy for the information environment that shapes how your company is understood. Not for public opinion itself, which is a step further on, but for the material opinion gets built from. That makes this reputation research as much as a discoverability exercise, and it is why it tends to land with the people who own what a company says and who says it, rather than with a search team.
The useful output is not a score. It is a list: which outlets are shaping the answers, which relationships are carrying weight, where your story is not landing, and which pages the agents reach for when someone is close to deciding.
Questions clients actually ask
"No one is lacking for GEO data. Where do we start, and what do we do with it?"
Pick one customer use case and study it closely rather than tracking several hundred prompts daily. Read the citations. Look at what is driving your presence, your prominence, and your absence. A daily tick up or down is mostly noise; a close read of one use case tells you what to do on Monday.
"How do we know which sites are influencing AI, and how do we prioritize them? We used to prioritize by size and influence. Now anybody can influence."
Correlate sources with outcomes. Look at which outlets, review sites, community platforms and owned pages appear when you are present and prominent, and which appear when competitors are. That gives you a targeting list rather than a scoreboard.
We would not advise chasing every review site that surfaces. It is hard to move and harder to hold. Look instead for owned or earned sources that can fill the same space on the topics where those sites are being cited.
In one read of a niche fragrance category, an online retailer's blog was the second most-cited source overall, with more than three times the citations of the largest beauty retailer. The largest marketplace was not cited once.
"How do we get to the right set of prompts?"
Garbage in, garbage out. Generic, self-referential prompts produce data you cannot act on. Model prompts have to be built from personas grounded in whatever first-party material you choose to share: search queries, ICPs, customer research. None of it is required, and none of it needs access to your systems.
Prompts are modeled, not observed. The companies behind these agents do not publish what people ask them. Anyone showing you "real prompts" at scale is showing a sample or a model, not the record. Say so, and design accordingly.
"Should we start with owned content or earned media?"
Directionally, owned tends to matter for high-intent questions and earned for research questions. But every category is different. Some are listicle-driven, some lean on forums and reference sites, some on traditional press, some on a brand's own library.
There is no reason to guess. Run a set of prompts and look at the split.
In one corporate-reputation read, every agent cited the company's own pages when asked about it by name. Across forty answers to unbranded category questions, those pages did not appear at all.
"How often should we re-measure?"
Monthly. Many dashboards run prompts daily, and daily variation is more noise than signal. We are not looking for basis points. We are looking for an indication that a strategy is working or that it needs to change.
"Can we go back and measure retroactively?"
Not really, and that is a genuine limitation. An agent's answer on a past date cannot be re-collected. A baseline measured now is the earliest comparison you will have, which is the argument for measuring before a launch or a campaign rather than after. From that day every run is stored in full, so later comparisons are exact.
"What is the quickest thing we can do?"
FAQ pages, written properly. They are still the fastest win for almost every client. They have to be informational and honest, in your own voice, and willing to answer the questions a buyer actually asks rather than the ones that flatter you. Add the first-party facts only you have.
Then the technical basics: clear page headings, schema markup, and, if you want to be cited, a robots file that does not block the retrieval crawlers. Those are separate from the training crawlers, and you can allow one and block the other. Blocking both is a legitimate choice for some publishers. It is just a choice worth making deliberately.
How a measurement works here
- Five AI platforms included, up to ten, and you choose which.
- 100 prompts, branded and unbranded, written with your team and approved before anything runs.
- Every answer kept in full, not summarized.
- Every citation captured and traced to the page it actually points at. Links that do not resolve are excluded and listed, never counted.
- Every distinct cited page fetched and checked for whether it mentions you. Refusals recorded rather than guessed at.
- Every headline number recomputed from the raw answers, independently of the system that produced it, and reviewed by Nathan before you see it. You keep the raw file.
What this cannot do
Nobody can steer these models. Not us, and not the people who build them. The honest description of this work is that we study them.
It is probabilistic, not deterministic. We do not promise visibility improvements, and we would be suspicious of anyone who does. Healthy skepticism about this category is warranted, and we share it.
What we do commit to is the read itself: what is being said, the sources behind it, a set of recommendations you can act on, and a re-measurement to see whether they worked.
Where to start
If you want to know what agents currently say about your brand, the fastest way to find out is to ask them. Bring one question you actually want answered and we will tell you whether a run would answer it.