Thank you for these questions — they go straight to the real limits of GEO.
1. Scalability and cost across an entire catalog
This is the question every merchant should ask first. Honest answer: no, this level of effort is not viable product by product on a 500-item catalog.
The right approach is prioritization. Identify the 5 to 10 flagship products — highest margin or highest volume potential — and apply the full method to those. For the rest, raise the bar compared to existing content, without chasing GEO perfection. It's a portfolio logic, not a uniformity one.
With generative AI today, going from 80 to 900 structured words takes 2-3 hours of work (brief, generation, revision, integration). On 10 priority products, that's one week. The value produced then holds for 12 to 18 months.
2. Indexing speed after a model update
Genuinely out of scope for this study, and I own that — training cycles for large models remain opaque. What we observe in practice: Perplexity (which crawls in real time) reacts within weeks. ChatGPT and Gemini, whose training data is frozen between versions, can take several months.
What is clear: well-structured content, cited by third-party sources, holds up better across model rotations than content optimized on a single signal. That's exactly why the ecosystem of mentions matters as much as the page itself.
3. Abuse, prompt engineering and data poisoning
A legitimate and underrated question. We're already seeing attempts to manipulate recommendations through content deliberately crafted to inject biased signals into training corpora.
Engines defend against this through source triangulation: an isolated signal carries little weight, a consistent signal across 15 independent sources carries much more. What also protects legitimate players: real credibility can't be bought in bulk. That said, we're only at the beginning of this arms race — it will be a major topic over the next 18 months.
4. Real sales attribution — the genuine limit of this study
You're right to call it out, it's the weakest angle of the case. The +75% in sales is real, but precise attribution between GEO, improved SEO, third-party blogs and forum activity can't be cleanly isolated without dedicated multi-touch tracking.
What I can state with confidence: the 64 visits from ChatGPT and Perplexity are directly traceable via UTM. The rest — the amplification effect on Google (position 8 → 5), the blog surge (+600%) — results from a combination that GEO triggered but didn't produce alone.
This is actually my implicit conclusion in the article: GEO and SEO are inseparable in their effects. And that's precisely why the ROI argument holds, even if pure attribution remains blurry.
I was expecting this answer. The main reason I asked the question is because there are companies with boring products. I don't see people creating that much content for a nail as an extreme example.
It is sad you didn't make the effort to track the buyer journey. An alternative would be a "where did you find us" question during the checkout flow coupled with some kind of incentive to answer it.
And that's precisely why the ROI argument holds, even if pure attribution remains blurry.
I don't doubt the effort had an effect. To me it looks like you were testing too many things at the same time. I would do a two step process, first SEO then GEO.
If SEO efforts can get products in LLM's there is no need for GEO improvements.
Thank you David for these sharp questions — they go right to the heart of what makes GEO genuinely difficult to scale.
On the cost of doing this for every product: You're absolutely right — it's not scalable manually. The realistic answer is to prioritize: identify your top 10–20 products that already have organic traction, and apply the full GEO treatment there first. For the rest, a lighter version (better structured data, a FAQ block, one supporting article) already moves the needle. In my workflow, I've started using AI-assisted content generation for the editorial layer, which reduces the 900-word overhaul from a full day's work to 2–3 hours. Still a real investment, but more manageable at catalog scale.
On model freshness and coverage: Perplexity is by far the fastest to pick up new content (days to weeks, as it actively crawls). ChatGPT with browsing follows within weeks. Base LLMs without real-time access won't update until their next training cycle — which is a known limitation I should have made more explicit in the article. GEO primarily targets retrieval-augmented models for now.
On prompt injection risks: Completely valid concern I didn't address. It's an emerging attack surface — structured data and FAQ blocks can theoretically be abused. I don't have a clean answer yet, but it's definitely worth a dedicated article.
On attribution of sales: Fair methodological pushback. The +75% in units is correlated, not proven causal. The honest answer is I don't have full journey tracking in this specific case. What I can confirm is that 41 referral visits came directly from Perplexity (measurable in GA4), and those converted at a rate comparable to organic. The AI-to-sale path exists — but you're right that the article overstates the causality.
These are exactly the questions a proper GEO study should address. Thanks for raising them publicly.
For further actions, you may consider blocking this person and/or reporting abuse
We're a place where coders share, stay up-to-date and grow their careers.
Thank you for these questions — they go straight to the real limits of GEO.
1. Scalability and cost across an entire catalog
This is the question every merchant should ask first. Honest answer: no, this level of effort is not viable product by product on a 500-item catalog.
The right approach is prioritization. Identify the 5 to 10 flagship products — highest margin or highest volume potential — and apply the full method to those. For the rest, raise the bar compared to existing content, without chasing GEO perfection. It's a portfolio logic, not a uniformity one.
With generative AI today, going from 80 to 900 structured words takes 2-3 hours of work (brief, generation, revision, integration). On 10 priority products, that's one week. The value produced then holds for 12 to 18 months.
2. Indexing speed after a model update
Genuinely out of scope for this study, and I own that — training cycles for large models remain opaque. What we observe in practice: Perplexity (which crawls in real time) reacts within weeks. ChatGPT and Gemini, whose training data is frozen between versions, can take several months.
What is clear: well-structured content, cited by third-party sources, holds up better across model rotations than content optimized on a single signal. That's exactly why the ecosystem of mentions matters as much as the page itself.
3. Abuse, prompt engineering and data poisoning
A legitimate and underrated question. We're already seeing attempts to manipulate recommendations through content deliberately crafted to inject biased signals into training corpora.
Engines defend against this through source triangulation: an isolated signal carries little weight, a consistent signal across 15 independent sources carries much more. What also protects legitimate players: real credibility can't be bought in bulk. That said, we're only at the beginning of this arms race — it will be a major topic over the next 18 months.
4. Real sales attribution — the genuine limit of this study
You're right to call it out, it's the weakest angle of the case. The +75% in sales is real, but precise attribution between GEO, improved SEO, third-party blogs and forum activity can't be cleanly isolated without dedicated multi-touch tracking.
What I can state with confidence: the 64 visits from ChatGPT and Perplexity are directly traceable via UTM. The rest — the amplification effect on Google (position 8 → 5), the blog surge (+600%) — results from a combination that GEO triggered but didn't produce alone.
This is actually my implicit conclusion in the article: GEO and SEO are inseparable in their effects. And that's precisely why the ROI argument holds, even if pure attribution remains blurry.
Thank you for answering my questions.
I was expecting this answer. The main reason I asked the question is because there are companies with boring products. I don't see people creating that much content for a nail as an extreme example.
It is sad you didn't make the effort to track the buyer journey. An alternative would be a "where did you find us" question during the checkout flow coupled with some kind of incentive to answer it.
I don't doubt the effort had an effect. To me it looks like you were testing too many things at the same time. I would do a two step process, first SEO then GEO.
If SEO efforts can get products in LLM's there is no need for GEO improvements.
Thank you David for these sharp questions — they go right to the heart of what makes GEO genuinely difficult to scale.
On the cost of doing this for every product: You're absolutely right — it's not scalable manually. The realistic answer is to prioritize: identify your top 10–20 products that already have organic traction, and apply the full GEO treatment there first. For the rest, a lighter version (better structured data, a FAQ block, one supporting article) already moves the needle. In my workflow, I've started using AI-assisted content generation for the editorial layer, which reduces the 900-word overhaul from a full day's work to 2–3 hours. Still a real investment, but more manageable at catalog scale.
On model freshness and coverage: Perplexity is by far the fastest to pick up new content (days to weeks, as it actively crawls). ChatGPT with browsing follows within weeks. Base LLMs without real-time access won't update until their next training cycle — which is a known limitation I should have made more explicit in the article. GEO primarily targets retrieval-augmented models for now.
On prompt injection risks: Completely valid concern I didn't address. It's an emerging attack surface — structured data and FAQ blocks can theoretically be abused. I don't have a clean answer yet, but it's definitely worth a dedicated article.
On attribution of sales: Fair methodological pushback. The +75% in units is correlated, not proven causal. The honest answer is I don't have full journey tracking in this specific case. What I can confirm is that 41 referral visits came directly from Perplexity (measurable in GA4), and those converted at a rate comparable to organic. The AI-to-sale path exists — but you're right that the article overstates the causality.
These are exactly the questions a proper GEO study should address. Thanks for raising them publicly.