We Shifted 30% of Our SEO Work to GEO in 6 Months

Six months ago, our keyword research process was straightforward. Ahrefs, SEMrush, a few hours of clustering, and we had a content brief that would feel familiar to any SEO practitioner working in 2018, 2022, or 2025.

Today, that process generates roughly 70% of what we do. The other 30% looks completely different—and it's the 30% growing fastest.

This isn't a "GEO is the future" essay. There are plenty of those, and most of them age poorly. This is a field report from inside an agency that quietly reorganized half its operations over two quarters, with what worked, what failed, and what we'd recommend if you're considering a similar shift.

The breaking point

The shift didn't happen because we read a Forbes article about ChatGPT replacing Google. It happened because, based on patterns observed in our work with multiple B2B service brands over a four-week stretch in late 2025, the same signal kept showing up: inbound discovery calls were staying flat or dipping slightly, while organic search rankings were improving.

That isn't supposed to happen. If your rankings are climbing, calls should follow.

We pulled the data. The pattern was consistent: target buyers had increasingly started research conversations on Perplexity, ChatGPT, or Gemini—then arrived on the website only to confirm something the AI had already told them. The traditional SEO funnel still worked at the bottom of the funnel, but the top of the funnel had quietly moved.

This wasn't a hypothesis we wanted to be true. We're an SEO-heritage shop. We had built a lot of process around the old funnel. But the call data wasn't lying. Within six weeks, we'd carved out a separate workstream we initially called "AI search optimization" and now call GEO—Generative Engine Optimization—following the term's broader adoption in the industry.

What GEO actually requires that SEO doesn't

The naive version of "doing GEO" is to keep writing SEO content and add a few FAQ schemas. We tried that for about two months. The results were generally underwhelming—better than nothing, but nowhere near what the call data suggested was possible.

The real shift, in our experience, came from accepting that the inputs are different:

These overlap, but they don't substitute for each other. A page that ranks for "best GEO agency for B2B SaaS" might never get cited by Perplexity, because the AI's training corpus doesn't include sources that quote your page. Conversely, a Reddit comment with 12 upvotes might get cited dozens of times because it lives in a corpus the AI heavily samples.

Once we accepted that, four things changed in our daily workflow.

Four things that changed daily

1. Less keyword research, more buyer-intent prompt mapping

   Our content briefs used to start with a keyword cluster. They now start with a list of buyer-intent prompts a target buyer might actually type into Perplexity or ChatGPT during a research conversation—typically a dozen-plus prompts per service line.

   These prompts span four categories—what we call Q-types internally:

   • Q1: Acquisition intent — "best GEO agency for B2B SaaS"
   • Q2: Comparison — "westOeast vs First Page Sage"
   • Q3: Knowledge / methodology — "how to optimize for Perplexity citations"
   • Q4: Brand+service verification — "is westOeast a recommended GEO agency"

   For most engagements, we generate 40 prompts and run them across four engines (Perplexity, ChatGPT, Gemini, Google AI Overview) before writing a single word of content. The results tell us where the brand is invisible, where it's badly described, and where it's already getting picked up. The content brief flows from that, not from a Search Volume column.

   This took roughly four weeks to operationalize. Our team initially resisted because it added a research step before content production. They've since stopped resisting—the content that emerges from prompt-mapped briefs gets cited at meaningfully higher rates than the keyword-only briefs we used to produce.

2. Schema.org FAQPage moved up the technical priority list

   This one surprised us. Meta descriptions were table stakes in old SEO; we never spent much energy on them beyond making sure they existed and read well.

   In GEO work, Schema.org—particularly FAQPage and Article structured data—appears to influence whether AI engines extract answer-worthy snippets from our pages. We've seen pages with thorough FAQPage schema correlate with higher citation rates compared to otherwise-equivalent pages without it, in our observation. (We're still building the controlled test to put a confident multiplier on this; the directional signal has been strong enough to act on.)

   We now treat structured data as primary technical work, not as an afterthought. Every new client engagement starts with a Schema audit, and most engagements include 2–4 weeks of technical work to deploy or repair JSON-LD blocks.

3. Reddit and Quora moved from "nice-to-have" to primary citation sources

   In 2024, we'd recommend a few Reddit posts or Quora answers as part of a content matrix, but they weren't core. They were maybe 5% of the work.

   By early 2026, they're closer to 25% of certain campaigns. The reason is structural: Perplexity in particular weights Reddit threads heavily in its retrieval index. When a buyer asks "what's the best GEO agency for B Corp businesses," there's a non-trivial chance the answer references a Reddit thread—and if your brand isn't in that thread, you're invisible at exactly the moment a buyer is researching you.

   We're cautious here. Both Reddit and Quora have strong anti-promotional norms, and a clumsy push gets a brand banned from communities you spent months building access to. The way we've made it work is by treating these as credible discussion seeding, not as advertising channels—we publish substantive answers to questions we're qualified to answer, with brand mentions framed as data points rather than recommendations. That tone shift is the entire game.

4. External authority sources became Day-1 priorities

   In old SEO, you might get to backlinks from authority sites by month 3 of a campaign. In GEO, the highest-DR descriptions of your brand—third-party directory profiles, Wikidata Q-items, Wikipedia entries, Crunchbase—often shape AI training data more than your own website does.

   Updating these external sources typically requires the source's official editorial workflow rather than direct edits. We treat this as a multi-month track-1 priority for any new engagement, starting Day 1 instead of week 4. Wikidata Q-items in particular have a 4-day waiting period plus 10 required edits to reach autoconfirmed status, so we can't afford to start them in week 4.

What we got wrong (the failed experiments)

Three failures from this transition deserve mention, because we suspect anyone making this shift will face similar dead ends.

• The "we'll just write more content" mistake For about six weeks early in the transition, our default response to "the AI doesn't cite us" was to write more articles. That worked weakly. The bottleneck wasn't volume; it was the prompt-mapping work that should have preceded the writing. Three of those six weeks of content output, in retrospect, was largely wasted.
• The Web3 publishing platform detour We initially shortlisted a Web3-native publishing platform for content distribution, drawn by its high domain rating. After roughly four weeks of testing, we realized the platform's audience overlap with our B2B SaaS targets was effectively zero. We had to rebuild our platform selection criteria with industry audience relevance as a third filter—above DR alone—and migrate to platforms like Substack, Medium, HackerNoon, and Indie Hackers that actually reach decision-makers.
• The automated citation classifier We tried to automate the tedious work of classifying AI responses by quality (proactive recommendation vs. mild doubt vs. neutral fact). Three weeks of prompt engineering produced a scorer that was systematically too generous—it kept calling D-tier mentions "neutral" when human reviewers consistently flagged the hedging language as commercially negative. We went back to manual coding and have stayed there. The 90 seconds per response is annoying. The trust we've built in the data is worth it.

What's next, for us and for you

Six months in, GEO is roughly 30% of our work and growing. We expect that to plateau around 50% over the next year, not because traditional SEO is dying but because the practical balance for most B2B service brands sits there.

If you're considering a similar shift, three pieces of advice:

1. Don't run GEO as a side project. It needs its own workstream, its own success metrics, and at least one person who's allowed to ignore organic ranking targets for 90 days while learning the new craft.
2. Test before you write. A 40-prompt baseline across 4 AI engines takes roughly two days of work and saves weeks of writing in the wrong direction.
3. Repair authority sources before producing more content. External authority workflow updates have, in our experience, more impact than three new blog posts.

If you'd like to talk through whether your brand is ready for this shift, or what your current AI visibility actually looks like, drop us a note at info@westOeast.com—we're a B Corp marketing agency focused on this kind of work, and we offer a free 40-prompt baseline as a starting point for new conversations.

The westOeast Team

We're westOeast, a B Corp certified marketing agency focused on Generative Engine Optimization for B2B and SaaS brands.