Foxcite | AI Search Intelligence

Search isn't dying. It's growing a second layer.

A buyer who used to type "best payroll software for restaurants" into Google and open five tabs now often asks ChatGPT, Perplexity, Gemini or Google's own AI Overview, and reads one synthesized answer. That answer names a few brands, cites a few pages, and frequently ends the search right there.

That layer is where discovery is moving, and it plays by rules that overlap with SEO without being the same. Answer Engine Optimization (AEO), also called Generative Engine Optimization (GEO), is the work of getting your clients named and cited inside those answers.

This post covers what has actually changed, how answer engines pick their sources, which SEO habits still pay off, and where an agency's effort goes furthest in 2026.

What the data says (and doesn't)

It's easy to overstate this shift, so let's start with the numbers we can stand behind.

Fewer clicks when an AI summary appears. In a Pew Research Center study of real browsing data from 900 U.S. adults, users clicked a traditional search result on 8% of visits when Google showed an AI summary, versus 15% when it didn't. Only 1% of visits included a click on a link inside the summary itself.
Lower click-through for top rankings. Ahrefs analyzed 300,000 keywords and found that position-one pages earned about 34.5% lower click-through rates on informational queries that showed an AI Overview. Ahrefs presents that as a correlation, not proof of cause, and so should you.
AI referrals are still small. Conductor's 2026 AEO/GEO benchmark, covering about 13,770 enterprise domains, put AI referrals at around 1.08% of site traffic on average, with most of it coming from ChatGPT. Organic search still dominates.

Put those together and you get the honest picture. AI engines don't send much traffic yet, but they are absorbing the click that used to go to your client's page. The influence happens in the answer, before anyone visits a site, and a referral number can't capture it. A buyer who reads "the top options are X, Y and Z" and then searches for Y by name shows up in your analytics as branded search, not as AI traffic.

That's why AEO needs its own measurement. If you judge it by referral sessions alone, you'll conclude it doesn't matter right up until your client's competitor owns every answer in the category.

How answer engines choose sources

You don't need to know model internals to do this work, but you do need a working model of the pipeline. For most answer engines with web access, it looks roughly like this:

Interpret the question. The engine rewrites the user's prompt into one or more search queries. A question like "which CRM is best for a 10-person real estate team" may fan out into several sub-queries about CRMs, real estate, and small teams.
Retrieve candidates. It pulls pages from a search index: its own, Bing's, Google's, or a mix, depending on the engine.
Select and extract. It picks the passages that answer the sub-questions most directly and most credibly.
Synthesize and cite. It writes an answer and attaches citations to some of the sources it used.

Two things follow from this.

First, retrieval still runs on search fundamentals. If a page isn't crawlable, indexed and relevant to the sub-queries, it never reaches the selection step. Most of what you already do in technical SEO is the price of admission.

Second, selection rewards extractability and credibility. Engines are looking for passages that answer cleanly and that they can attribute to a trustworthy source. A 3,000-word page whose real answer is buried in paragraph nine competes poorly with a page that answers in its first two sentences.

The academic work supports the second point. The GEO paper from Aggarwal and colleagues (KDD 2024) tested content changes against generative engines and found that adding citations, quotations and statistics could raise a source's visibility by up to 40% in their experiments. "Up to" carries a lot of weight in that sentence, and the authors note the effect varies by domain. The direction is still clear: evidence-dense content gets picked more often than vague content.

What carries over from SEO

Plenty of what agencies already know still works. Some of it matters more than before.

Technical health. Crawlability, indexation, fast server responses and clean rendering decide whether a page reaches the candidate set at all. Check your client's robots rules for AI crawlers too. Some sites block GPTBot, ClaudeBot or PerplexityBot by accident through a blanket rule copied from a template.

Topical authority. Engines retrieve from sites that have depth on a subject. One great page on a thin site rarely beats a solid page on a site that covers the topic thoroughly.

Links and mentions. Third-party coverage matters, perhaps more than before, because answer engines lean heavily on review sites, comparison articles, forums and editorial lists when they recommend brands. For many commercial questions, the page that gets cited isn't your client's site at all. It's a "best X for Y" roundup that mentions them.

Structured data. Schema helps engines understand entities (who the organization is, who wrote the page, what the product is), but keep expectations realistic. Google's own guidance for AI features says there is no special markup required to appear in them, and that structured data should match the visible content. Treat schema as clarity, not as a ranking switch.

What's genuinely new

The unit of competition is the answer, not the page. In classic SEO, ten blue links meant ten chances. In an AI answer, there may be three to six brand mentions and a handful of citations. Being mentioned in second place in an answer is a different outcome from ranking second.

Answers vary. Ask the same question twice and you may get different brands. Ask a different engine and you'll often get a different answer again. Visibility is a probability, not a position. You measure it by running the same prompts repeatedly across engines and tracking how often your client appears.

Mentions and citations are different wins. An engine can mention a brand without linking to it, or cite a page without naming the brand in the text. Both matter, for different reasons. A mention shapes the shortlist. A citation sends some traffic and signals that the engine trusts the page as a source.

Prompts are longer and more specific. People talk to AI assistants in full sentences with context: budget, location, team size, constraints. That rewards content that addresses specific situations, not just head terms.

Myths worth retiring

"Just add an llms.txt file." It's a reasonable proposal and cheap to add, but no major answer engine has confirmed it uses llms.txt when choosing sources. Don't sell it as a strategy.

"AEO replaces SEO." It sits on top of SEO. Weak fundamentals sink AEO work before it starts.

"You can buy your way into ChatGPT." Not through the organic answer. There are paid placements in some AI products, and more are coming, but they're labeled as ads and are a separate channel.

"Write for the bot." Content stuffed with question-shaped headings and robotic answer blocks reads badly and earns no links or mentions. Write clearly for people, and structure it so a machine can lift a passage without losing the meaning.

Where to put the effort

If you're advising a client this year, this is the order I'd work in.

Measure first. Build a set of 25–100 prompts that real buyers would ask, across the main engines. Record which brands get mentioned, which pages get cited, and where your client is missing. Without a baseline you can't show progress, and you'll guess at priorities.
Fix the retrieval blockers. Crawler access, indexation, rendering, and the obvious technical gaps on the pages that should answer those prompts.
Make the key pages answerable. Lead each section with a direct answer. Add the specifics engines look for: prices, ranges, steps, comparisons, named sources. Keep facts consistent across the site so the engine doesn't find three different answers to the same question.
Earn third-party presence. Look at which outside pages get cited for your client's prompts. Those reviews, directories, roundups and forum threads are your outreach list.
Re-measure on a schedule. Answers drift as models and indexes update. Monthly is a sensible minimum, and weekly if the category moves fast.

A short example

Say a regional HVAC company wants to show up when people ask "who should I call for a heat pump installation in Denver?"

The baseline shows it's mentioned in about one in ten answers on Perplexity, never on ChatGPT, and that ChatGPT keeps citing a local "best HVAC companies" article that doesn't list it. Its own heat pump page talks about "comfort solutions" and never states a service area, a price range or a rebate amount.

The fixes almost write themselves. Rewrite the page to state the service area, typical installed cost ranges and the current rebates, with sources. Add a clear FAQ that answers the questions people actually ask. Contact the publisher of the roundup with a case for inclusion. Then re-run the same prompts and see whether the numbers move.

None of that is exotic. It's ordinary SEO judgment pointed at a new target, with measurement that matches how the target behaves.

FAQ

Is AEO different from GEO? In practice, no. Both describe optimizing for AI-generated answers. AEO is common in agency circles and GEO in academic and tooling circles. Pick one and explain it to clients once.

How long does AEO take to show results? Engines with live web retrieval, like Perplexity and Google's AI features, can reflect changes in days to weeks once pages are recrawled. Answers that lean on a model's training data change slowly, on the model's release cycle. Set client expectations in months, and report progress monthly.

Does ranking on Google still matter? Yes. Several engines retrieve from search indexes, and Google's own AI features draw from its index. Strong rankings make it more likely your page is a candidate when the engine assembles an answer.

Which engines should we track? At minimum ChatGPT, Google AI Overviews, Perplexity and Gemini. Add Claude, Copilot or Grok if the client's audience uses them. The answers differ enough that tracking only one gives a misleading picture.

Can we guarantee a client gets cited? No, and you shouldn't promise it. You can promise a measured baseline, a clear plan and regular reporting on mention and citation rates.

The bottom line

The AI discovery layer rewards the same things good SEO always rewarded: being findable, being clear, and being trusted by other sources. What changes is the scoreboard. You can't see AEO results in a rank tracker, so the agencies that win will be the ones that measure answers directly and act on what they find.

Foxcite was built for that measurement. It runs your clients' prompts across ChatGPT, Claude, Perplexity, Gemini, Grok and Google AI Overviews, and shows where they're mentioned, where they're cited, who is taking their place, and what to do about it.

Sources

The Future of AEO: Winning the AI Discovery Layer in 2026