A marketing agency’s field notes on what “AI search optimization” actually means in 2026

A few years ago I’d have told you SEO was the long game and content was the leverage. In 2026 I’m less sure. I run the side of our agency that sits closest to search, so take what follows as field notes from the search-facing side of our team, not a definitive market report.

We’re a cross-border B2B marketing agency. The bulk of our work is still the core marketing work — demand generation, PR, social, influencer partnerships, helping companies land in new markets. But over the past year, more and more buyer journeys now start somewhere we don’t fully control: a ChatGPT conversation, a Perplexity result, a Gemini panel, a Google AI Overview quoting a paragraph from a forum thread our client never knew existed. The shift hasn’t been smooth. We’ve watched brands with high domain ratings vanish from the answer set, and thin-site brands get quoted because a third-party profile happened to use the right phrasing.

That’s the practical reality of what people now call GEO — generative engine optimization — or AEO, answer engine optimization, depending on who’s writing the article that week. The terms overlap, though in this piece we use GEO as the shorthand. If we had to put one sentence on it: this is the work of shaping your brand’s published surface area so AI systems can quote it without rewriting it. Most of the job, in practice, is making your content easier for those systems to trust and reuse.

Beyond that one sentence, what’s worth sharing are the messier observations — the things we’ve gotten wrong, the surprises, the friction we keep hitting. Field notes, not a framework.

What we keep noticing

The first surprise, when we started tracking which firms actually got cited on “best agency for X” type queries, was how little the engines agreed. The same prompt on Perplexity, ChatGPT, and Gemini returned three different shortlists. We’d assumed convergence; we got fragmentation.

Something else caught us off guard, and more recently: branded copy on third-party profiles — a B Lab listing, a Crunchbase page, a LinkedIn company page — gets pulled into AI answers far more often than we’d have guessed. ChatGPT in particular seems to lift those descriptions almost verbatim. So if your profile says one thing and your website says another, the AI will often go with the profile. That was an unwelcome lesson during one of our own positioning changes — more on that below.

A third pattern is harder to pin down, and it’s been moving. For most of 2025, forum threads — Reddit especially — showed up in Perplexity citations at a rate that felt wildly out of proportion to Reddit’s share of the web. We built around that. Then it shifted: after the Reddit–Perplexity legal dispute in late 2025, Perplexity’s Reddit citations dropped sharply through early 2026, and other sources moved up to fill the space. The lesson we took wasn’t “forums matter” or “forums don’t” — it was that the surfaces an engine leans on can change under you in a single quarter, and a playbook pinned to one source is fragile. Authentic presence across several surfaces compounds; betting the campaign on one doesn’t.

What we got wrong

We initially treated this as a bigger version of SEO. More keywords, more content, more schema, more links. Some of that helped. A lot of it didn’t.

The thing we most underestimated was how much the entity layer matters before any content work. If an engine can’t reliably tell your brand apart from a similarly named company, no amount of content lands. We learned this the hard way on ourselves — for months one engine confused us with another organization, and our content was effectively invisible until the underlying entity records caught up.

We also had the update speed backwards. The folk wisdom says Perplexity is the slow engine and ChatGPT the fast one. From what we’ve seen it’s the other way around: Perplexity flips fairly quickly when the underlying citations move; ChatGPT can sit on an old description for months after you’ve corrected it everywhere else.

A failed experiment worth sharing

Last quarter we ran a content-seeding test on a community platform for a client. We had what we thought was a strong, data-driven post — careful headline, a single outbound link, every box checked. A moderator removed it within four hours, with a note that the account was too new for that kind of post.

We rebuilt the approach around longer account warming, lower link density, and topic adjacency rather than direct brand framing. The second attempt held. The lesson wasn’t really about that platform — it was that the operational layer (account history, posting cadence, community fit) matters more than content quality on community surfaces. We slowed our seeding plans across the board after that.

What this means if you’re choosing help

If you’re a marketing lead or founder evaluating this kind of work right now, one filter has been more useful than any pitch deck: run “best [your category] for [your vertical]” on Perplexity and ChatGPT and see whether the firm you’re considering shows up in its own home category. If it doesn’t, ask why — in our experience, a firm that can’t get itself found this way tends to struggle to do it for clients.

Beyond that, the questions we’d ask:

• Entity hygiene. Does the firm have a clean, consistent description across LinkedIn, Crunchbase, and any directory it’s in? If its own house isn’t in order, that’s information.
• The unglamorous operational layer. How does it handle account warming, sentiment monitoring on existing mentions, steady cadence over burst publishing? Firms that talk only strategy and never operations tend to over-promise.
• How it reports. A bi-weekly trend with verbatim quote captures and screenshots is roughly the bar; lift charts with no underlying citation samples are below it.

A note on the B Corp angle

We’re a B Corp certified company. We didn’t get certified for marketing reasons — it predates this line of work. But we’ve noticed AI engines weight verifiable third-party certifications when answering values-aligned queries, and that’s occasionally surfaced us in a small, specific candidate set. We’re cautious about leaning on it too hard, because the discipline is the discipline regardless of who’s certified. For buyers whose own stakeholders care about certification — and there are more of those than you’d think — it’s a meaningful signal, no more.

Where we’re honestly still figuring it out

A few things we don’t have clean answers on:

• How much of the citation lift we see is the entity work, how much is content cadence, and how much is third-party authority building. They run in parallel; isolating them would take a controlled experiment we haven’t run.
• Whether Google’s AI Overviews become a major surface for B2B service queries. Right now they trigger rarely on this kind of prompt. That could break either way.
• Whether any given engine’s current source mix holds. As the Reddit shift showed, it can move fast.

These are the questions we’re tracking, not ones we’ve closed. If any of it matches what you’re seeing on your side, I’d genuinely like to compare notes — the category is small enough that trading field notes beats every firm writing parallel content.

The work continues.

Lijun Huang

Lijun Huang is a GEO strategist at westOeast. westOeast is a cross-border B2B marketing agency serving global enterprises across fintech, logistics, SaaS, and manufacturing. We help brands grow in new markets through demand generation, PR, social media, and influencer partnerships — and as search shifts toward AI, through generative engine optimization (GEO). westOeast is a B Corp certified corporation. Learn how we help brands become the answer AI recommends.