The Honest View. How Do LLMs Compare Against Other Performance Channels?

A paper in Marketing Science (Frontiers: ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels?) is the first large-scale, independent look at how AI referral traffic performs against traditional channels.

It is worth reading because it offers a different view from all the case studies doing the rounds, which I found very interesting and topical for our clients.

What the paper analysed:

• 973 e-commerce websites, roughly $20 billion in combined annual revenue

• 12 months of first-party Google Analytics data, August 2024 to July 2025

• More than 50,000 transactions from ChatGPT referrals, set against 164 million from traditional channels

• Financial metrics measured: conversion rate, average order value, revenue per session

• Engagement measured: bounce rate, session duration, pages per visit

• ChatGPT only. It drives over 90% of AI referral traffic, so the authors focused there. Perplexity, Gemini and Copilot trail in low single digits

The authors are refreshingly upfront about how new this channel is and that we don't have it mapped out yet. ChatGPT only added outgoing links in August 2024! Throughout the study window, those links were fully organic, not paid. That has since changed with OpenAI rolling out advertising and shopping features inside ChatGPT. So treat everything below as a snapshot of a channel that is roughly a year old and moving fast!

Picture the kind of query your future guest or buyer is now typing into ChatGPT:

Planning a 4-night anniversary stay in London. I want a spa and, ideally, Michelin-recognised restaurants nearby. Where should we stay?

Where is the best place to live in East London? It needs a state-of-the-art gym, great co-working, lounges I can host friends in, ideally some outdoor space, even better with a paddle court, and it must be close to a tube station so I can reach central London in under 30 minutes.

No keyword search handles that well. An LLM reads every constraint, weighs the trade-offs, and hands back a shortlist with reasons. For high-consideration, multi-variable purchases, that is a materially better experience - no matter how much we know AI hallucinates or serves out of date information, it a trade off us consumers are willing to take.  We’re lazy creatures, and curious to see if it will point us in a direction we hadn’t thought of before. To top it all off, at the end of that answer sometimes sits outgoing links to the brands the model chose to feature.

A lot of agency blogs are quietly misleading. You will have seen figures like 6.7% conversion for AI referrals versus 3.9% for organic search, or claims that a session from ChatGPT is worth 4.4 times an organic one. Those come from early industry studies, and the paper cites them as exactly that. Its own controlled analysis, which adjusts for website differences and thin data, finds something far more sober:

• AI referral traffic is under 0.2% of all sessions, about 200 times smaller than Google's organic search

• On conversion rate and revenue per session, it sits below every traditional channel except paid social

• Organic search converts roughly 13% higher than AI referrals in the broad data

So no, on today's aggregate numbers, AI is not out-converting Google. The rigorous independent report says the hype is running ahead of the data.

On that 4.4x figure specifically: revenue per session is conversion rate multiplied by order value, and the early multiples are inflated by who is using these tools. Early adopters are self-selected, high-intent, digitally fluent researchers who arrive having already done their comparison inside the chat. Impressive numbers, unrepresentative sample.

That all said, There’s a plot twist for Effect’s clients. The paper finds that product complexity changes everything. The more considered and information-heavy the purchase, the better AI referrals perform. In high-complexity categories:

• AI referral conversion exceeds paid social, referral, email, organic search and direct

• Websites in these categories see 4.6 times the AI traffic share of simple-product sites

• Revenue per session moves up towards organic and paid search levels

And it skews to exactly the audience luxury brands court. Sites with younger, more technophile visitors show 3.8 to 5.5 times the AI traffic share. Premium travel and property audiences are more digitally fluent than their persona’s allow you to see. 

Luxury property and hospitality are textbook high-complexity, high-consideration, high-value purchases. This is the one corner of the market where the paper's data works firmly in your favour. Add luxury order values into the mix and even small volumes become commercially real. A single qualified enquiry on a five-figure booking is not a rounding error. This is the "small but mighty" case. 

For day-to-day, low-consideration retail, this channel is not there yet - maybe that will change when we can start transacting through our AI.  But, for luxury and premium brands, this is where we need to be playing now.

The study uses last-click attribution, the industry standard, but it flatters lower-funnel channels and undercredits discovery ones. Paid social doing research-phase steering on a considered purchase rarely gets the last click, and so its role is understated in this report. The same logic runs in AI's favour: a buyer who researches a place to live in ChatGPT, then converts weeks later through branded search or direct, hands all the credit to the last touch. So what do I take from the report?

1. AI's true contribution to the luxury purchase journey is almost certainly larger than these headline numbers suggest. 

2. Paid social is not being replaced by AI. I believe they have a symbiotic relationship. What you see in social will influence your prompt. What you are recommended in an LLM, you’ll search on Meta or TikTok. The funnel or as Google call it the 'Messy Middle' is still there!

The paper notes, almost in passing, that as this channel grows, website design will need to adapt. It does not say how.

If you want AI systems to find you, understand you, and put you on the shortlist, this is what needs to be done, and it is the core of our Hybrid Engine Optimisation approach:

Let the crawlers in. Many AI bots cannot render JavaScript, so server-side rendering is now a baseline, not a nice-to-have. 

Serve clean, machine-readable content. Industry-specific schema (Real Estate, Local Business, Product), short self-contained sections, descriptive headings, and tools like Cloudflare's Markdown for Agents that hand AI clean content instead of bloated HTML.

Publish files that speak to AI directly. llms.txt and ai.txt at the root of your domain tell AI systems who you are, what you offer, and how to represent you.

Write for extraction. Lead with the answer, use FAQ and comparison formats, and target the sub-questions inside a complex query rather than generic labels like "Our Services". I stress, however, that you must write for the human and not the bot; don’t overthink this. Write for your customer but follow the template we give you. 

Feed the recency bias. AI strongly favours fresh content, so key pages need genuine "last updated" signals and a quarterly refresh, not a one-and-done launch.

Build authority off your own site. AI trusts third parties more than self-promotion. Reviews, editorial coverage, consistent entity signals across Google Business Profile and Wikidata, and getting mentioned in the sources AI already cites all shape whether you get recommended. 

AI systems disproportionately cite original research and proprietary data, precisely the ethically sourced first-party data that sits at the centre of how we build. The brands that own genuine insight are the ones AI quotes.

The paper's window closes just as the channel is getting more interesting. Platforms are already moving on the biggest current limitation: that AI adds less value for simple purchases, with instant checkout and agentic shopping tools built for routine buying. Retailer-embedded assistants like Amazon's Rufus will compete with the general models by keeping customers in familiar interfaces.

And then there is advertising. AI referral traffic has been unpaid and organic, which makes it attractive to anyone watching return on ad spend. Paid formats inside ChatGPT change that calculus. Depending on how they are implemented, sponsored placements could smooth simple purchases or shift the balance between honest guidance and promotion in ways that affect trust. This is live, it is early, and it is exactly what we at Effect are testing now.

AI referrals are serving high-intent, digitally fluent customers making complex decisions. For luxury property and premium hospitality, it is a reason to move, because your customers, your price points, and your considered purchase journeys are precisely the conditions under which this channel performs. Early movers build advantages that compound, and the brands that get discoverable now will be on the shortlist when the volume arrives.

The original report entitled Frontiers: ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels? can be found here: https://pubsonline.informs.org/doi/full/10.1287/mksc.2025.0489?af=R
Published on 21 Apr 2026

Effect Digital builds end-to-end digital ecosystems for premium property and travel brands, with ethical first-party data at the core.

👋🏻 I’m launching a series about Hybrid Engine Optimisation (think GEO, AIO, LLMS) on my LinkedIn.  Follow me here: https://www.linkedin.com/in/natasharileyeffect/ 👈🏻