Full disclosure: This post contains affiliate links. If you make a purchase through these links, 99signals may earn a commission at no additional cost to you.
For the past two years, generative engine optimization (GEO) has been discussed as though it were additive. You optimize for AI citability, you gain visibility in ChatGPT and AI Overviews, and your existing organic performance carries on unaffected.
An A/B test run by the travel booking platform Omio, using SearchPilot’s testing infrastructure, provides the first controlled evidence that this assumption doesn’t always hold.
Omio ran two separate GEO experiments. The first added brand USP modules to pages, which increased LLM traffic by 18% with no negative effect on Google organic performance. A clean win.
The second test added structured “key takeaways” content, a tactic that appears in nearly every GEO playbook published in the last eighteen months. It performed positively for LLM-driven traffic. It also indicated it would likely reduce Google organic sessions by 6.5%.
Omio declined to roll it out and developed follow-up iterations instead.
Why the Math Rarely Favors GEO
The 6.5% figure deserves context, because in isolation it sounds survivable.
For most sites, Google organic still sends substantially more traffic than ChatGPT, Perplexity, and every other AI referral source combined. A percentage gain in LLM referrals is being measured against a small base. A percentage loss in Google organic is being measured against the primary revenue channel.
When those two numbers are compared in absolute terms rather than percentage terms, even a modest organic decline can erase a large proportional AI gain. The LLM traffic increase looks impressive in a dashboard. The organic decline is quieter, larger, and shows up in revenue.
There’s a conversion quality dimension as well. An AI mention is not equivalent to a click from someone actively comparing prices with intent to purchase. The two channels don’t convert at comparable rates, which widens the gap between the headline metric and the business outcome.
The Detection Problem
The more uncomfortable finding here isn’t the tradeoff itself. It’s that Omio only caught it because they were running a controlled experiment that measured both channels simultaneously.
Most organizations don’t have that infrastructure.
Without simultaneous measurement across organic and LLM channels, a GEO content change that erodes organic performance produces a signal that’s almost impossible to isolate. Organic traffic declines gradually. Teams attribute it to a core update, seasonality, competitive movement, or general AI-driven click erosion, all of which are plausible and all of which are happening independently.
Meanwhile the AI citation metrics are improving, which reinforces the belief that the GEO program is working.
The result is a change that looks successful on the metric being watched while quietly degrading the metric that funds the department. That pattern can persist for months before anyone connects it to the content change that caused it.
Not All GEO Tactics Carry the Same Risk
The most actionable finding from the Omio tests is that the two experiments produced different outcomes.
Brand USP modules increased LLM traffic without harming organic. Key takeaways bullets increased LLM traffic at an organic cost. Same site, same testing methodology, opposite risk profiles.
That distinction matters because it undermines the premise of blanket GEO playbooks. A tactic that’s genuinely additive for one page type or one site may be net-negative on another, and there’s no way to know which without testing.
The plausible mechanism is content simplification. Summary-style modules can make a page easier for a language model to parse, retrieve, and cite. The same simplification can reduce the depth signals Google’s ranking systems reward, or shift how users interact with the page in ways that feed back into rankings.
SearchPilot’s own framing is worth adopting: external GEO recommendations should be treated as hypotheses rather than instructions.
What This Means for How You Measure
The practical standard emerging from this is net business impact across channels, not isolated citation counts.
That requires three things most teams don’t currently have in place.
Simultaneous visibility across both surfaces. You need to see Google ranking movement and AI citation movement against the same timeline, for the same pages. Measuring them in separate tools on separate reporting cycles makes correlation nearly impossible to spot.
A clean before-and-after baseline. Before shipping any GEO content change, you need documented keyword positions and traffic estimates for the affected page set, so a decline has something to be measured against.
Whole-site traffic context. Keyword-level data tells you positions moved. It doesn’t tell you whether the AI referral gain exceeded the organic loss in absolute sessions. That calculation requires traffic-source-level data.
How Semrush One Fits This Problem
The cannibalization problem is fundamentally a measurement problem, and specifically a problem of measuring two channels that most tooling treats as separate disciplines.
This is the kind of gap
Position Tracking gives you the organic-side signal. Daily keyword-level ranking data means that when you ship a GEO content change, you can watch the affected keyword set for position movement in the days and weeks that follow, rather than discovering a decline in a quarterly review.
AI Search Tracking gives you the other half. Brand mentions and citations across ChatGPT, Perplexity, Gemini, and Google AI Overviews provide the LLM-side data you need for comparison. Seeing both trend lines in the same platform is what makes the tradeoff visible, which is precisely what Omio’s testing setup provided and most teams lack.
Organic Research establishes the baseline. Analyzing organic positions, traffic estimates, and keyword portfolios before a GEO experiment gives you the pre-change benchmark. Without it, any post-change decline is unfalsifiable.
Site Audit helps isolate causation. When organic performance drops after a content change, the immediate question is whether the content change caused it or something technical did. Site Audit surfaces the on-page and technical issues that might be interacting with your GEO modifications, which narrows the diagnosis considerably.
Topic Research reduces the tradeoff risk upfront. The Omio result suggests the danger lies in simplification, stripping depth to gain parseability. Topic Research surfaces semantically related subtopics and content gaps, which helps you add the structure LLMs retrieve well without removing the depth Google rewards. Building both signals into a page is a better strategy than trading one for the other.
Traffic Analytics answers the net-impact question. Estimating total visits by traffic source, including referral, is what lets you compare AI referral gains against organic session declines at the whole-site level. This is the calculation that determines whether a GEO experiment was actually net-positive, and it’s the one keyword-level tools can’t perform.
The reason a bundled platform matters here rather than a stack of point solutions is that cannibalization is only detectable in the comparison. Tracking rankings in one tool and AI citations in another means the two datasets live on different timelines with different refresh cycles, and the relationship between them stays invisible until it shows up in revenue.
Semrush One offers a free 7-day trial with access to 55+ tools across the SEO and AI Visibility toolkits, which is enough time to establish a baseline before your next GEO change goes live.
The Bottom Line
The GEO hype cycle has entered its correction phase. Brands that deprioritized traditional search optimization in favor of AI visibility six months ago are now spending to recover organic traffic they gave up.
The Omio tests suggest the correction isn’t an argument against GEO. One of the two experiments was a clean 18% gain with no downside. The argument is against shipping GEO tactics on the assumption that they’re free.
Some are. Some aren’t. The only way to tell them apart is to measure both channels at once, before the change goes live rather than after the traffic is gone.
Related Articles