From Visibility to Action: What AEO Really Means Now

I've spent 18 years in SEO, and I've never seen the ground move this fast. AI answers are now the single biggest source of influence on what people buy, and there's no paid way into them. You can't bid on a recommendation. You either earn it or you don't show up.

The path has collapsed into three steps: get found, get recommended, convert. Meanwhile, the crawlers doing the finding take far more than they send back in clicks (Cloudflare's crawl-to-referral data makes that painfully clear). Ranking alone no longer pays the bills.

This is how I think about AEO today: where to start, what to measure, what to do about it, and what comes next as agents start to act.

Start with the business goal, not the prompt list

Most teams start AEO by dumping a few hundred prompts into a tracker. The better sequence runs the other way: business goals → AI TAM → prompt index → actions.

The goal decides what to track. And it isn't always "more mentions." Different companies use AI visibility data in very different ways, and the list keeps growing. These are some examples, not the full picture:

• NPS and brand sentiment
• Product feedback
• Competitor benchmarking
• Industry and market research
• Sales insights
• Support (what AI tells customers when they hit a problem)

Once the goal is clear, size the market. An AI market index starts from five inputs: your website, your product or service (positioning, messaging, features and so on), your category, your audience (ICP, personas, segments) and your competitors. From those it maps topics, search volume, market demand and the number of prompts in play.

How to define a prompt

A useful prompt is persona + intent + your own content + real search demand. Intent covers comparisons, recommendations, pricing, alternatives and so on.

There's a catch. AI answers vary with demographics, context, user history and behavior. So any prompt set is a sample, not the whole picture. The blind spots are real: the prompts you didn't think of, how your brand shows up inside them, and the fact that you're only ever seeing a small fraction.

Three pillars: mentioned, accurate, visible

AI visibility breaks down into three questions:

1. Mentioned. Are you recommended? Are you listed as an alternative? How are you positioned, and who does the AI call "the best"?
2. Accurate. When AI talks about your brand and product, is it right?
3. Visible. How often you appear overall. Useful, but right now there are too many signals competing to define it.

Recommendation share beats raw mentions

A mention isn't a win if the AI says "avoid this one." The metric worth tracking is recommendation share: mentions plus a breakdown of how the LLM frames you.

Mention tier	What it means
Championed	The AI actively recommends you as the answer
Shortlisted	You're one of a few options it suggests
Neutral	Mentioned without a clear stance
Mixed	Positives and negatives in the same breath
Discouraged	The AI steers people away from you
Not mentioned	You're absent from the answer

The goal is to move prompts up this ladder, toward championed. Visibility plus being recommended is the foundation everything else sits on.

Use the tier to decide what to do next

The ladder isn't only a score. Once you know which tier each prompt sits in, you know where to spend your effort, because each tier usually points to a different kind of work:

• Not mentioned. Start by finding out why. If you have no page that answers the question, create one. If you have one that is thin or outdated, optimize or refresh it. If the sources the AI cites never talk about you, that calls for a third-party mention campaign: reviews, comparisons, lists and communities. If the answer leans on video, a video may be the missing piece.
• Discouraged. Find the source of the negative framing. It may be an outdated page, a review, a competitor comparison or a real product gap. Fix what's inaccurate, answer what's fair, and work on the content that shapes the opinion.
• Mixed. The AI sees both sides of you. Refresh your pages to make the strengths clearer, address the objections that show up in the same breath, and add proof such as case studies, data and reviews.
• Neutral. You're named but not backed. Give the AI a reason to recommend you: clear positioning, specific use cases, comparison content and evidence that you're good at the thing the prompt asks about.
• Shortlisted. You're one of a few, so the work is winning head to head. Build "X vs Y" and "best for" pages, and earn placement in the third-party lists where you aren't yet in the first few.
• Championed. Protect it. Keep the pages fresh, keep an eye on the sources that support the answer, and extend what works to neighboring prompts.

That's what turns a list of prompts into a to-do list ranked by how much each action is likely to move you up the ladder.

From insight to action

The real blocker in AI search isn't data. It's overwhelm: the list of possible actions keeps growing, and there's no easy line from insight to impact. So start with the outcome you want (being championed) and work backward through the citation systems that feed it.

That system has two halves, and they aren't equal. What you publish and fix on your own site is one. What others say about you is the other, and it may well be the bigger one.

An AirOps study from October 2025 looked at 21,311 brand mentions across more than 500 commercial-intent, discovery-stage queries in GPT-5, Claude Sonnet 4.5 and Perplexity Sonar. It found that 85% of brand mentions came from external domains, and only 13.2% came directly from the brand's own domain. Brands were about 6.5 times more likely to be mentioned through a third-party source than through their own site.

A few more findings from the same study are worth acting on:

• Listicles, comparisons and reviews do the heavy lifting. Nearly 90% of third-party mentions came from that kind of content, and around 80% of the brands in it appeared among the first three companies discussed. Position matters as much as presence.
• Your own pages still matter for detail. Product pages made up 19.3% of first-party mentions and homepages 7.1%. The study suggests your site carries the most weight when someone asks about a specific product, rather than for suggestions of vendors.
• Each model has its own favorites. About 68% of brands appeared in only one of the three platforms, so being visible in one tells you little about the others.

The caveats: it studied top-of-funnel discovery queries, not the whole buyer journey, and AirOps sells AI visibility software, so read its recommendations with that in mind. The direction is still clear. If you only fix your own site, you're working on the smaller half. Earning a place in the comparisons, reviews and lists that AI tools pull from is where outreach comes in.

Every action should be:

• Prioritized. Fixes ranked by likely impact.
• Explained. What to change, and why.
• Grounded. Tied to your own site and your AI market index.
• Specific. "Create this page," "update this one with photos," "refresh this description," mapped to real URLs and prompts.

Briefs that carry context

The content workflow looks like this: set targets, review the comparison pages, refine the briefs, then apply a custom style. A good brief specifies content type, persona, journey stage and suggestions. Then it layers in what AI can't invent on its own:

• Notes from subject-matter experts
• Original research and industry data
• Product specs

Finally it checks intent and topical coverage.

Build the context layer across departments

Behind all of it sits a context layer most teams skip. AI can write fast, but it can only write from what it knows, and what it needs to know doesn't live in the SEO team. It's spread across the company, so building it takes input from several departments:

• Product marketing. ICP and personas, positioning, messaging, differentiators and competitor benchmarks.
• Sales. Call transcripts, objections, the questions buyers really ask and the words they use to ask them.
• Support and customer success. Customer feedback, tickets, recurring problems and how customers describe them.
• Product and engineering. Specs, features and limits, so every claim is true and current.
• Brand and content. Voice and tone, examples of good human writing, words and sentences to avoid, copywriting skills, and writing and structure rules for each page type.
• Legal and compliance. The claims you can and can't make.
• SEO and AEO. Update rules, intent and coverage checks, and the prompts and citations a page should be built around.

Examples of good human writing do more work than rules alone. Showing the model what you want, and what you never want to see, beats describing it.

Without this layer you get fluent, generic content that sounds like every competitor. With it, a draft carries your positioning, your buyers' own language and facts that are true. Treat the layer as a living library: give each part an owner, keep it in one place the whole team can use, and refresh it when your product, your messaging or your competitors change.

Next: agents that act, not just answer

Once you're visible and recommended, agents start taking actions on the user's behalf. The chain looks like this: prompt → MCP / WebMCP → tool call → API → transaction. The recommendation becomes the purchase, with no website visit in between.

That raises the obvious question: do you need to change your site, or build a separate one for LLMs? My answer: you mostly don't need a separate site, but a few technical layers matter more than before:

Layer	Why it matters
Rendering	Your site has to be rendered so search engines and LLMs can read it, not only after a browser runs JavaScript
llms.txt	Not useful for content discovery; possibly useful for agent actions
Schema	Structured facts machines can trust
MCP tools	Let agents act on your product directly
API endpoints	How an agent actually places the order or takes the payment
Markdown files	Clean, parseable versions of key content

Underneath it all, the fundamentals haven't changed: great content, a healthy technical site, and earned authority. Optimize onsite (how you rank, what LLMs say about you) and offsite (YouTube transcripts, reviews, communities). And treat it as a practice, not a project.