Good morning.

Ten Shopify stores handed an agency two years of raw analytics, and the best-converting traffic on any of those sites turned out to be the sliver arriving from AI assistants, worth about a tenth of one percent of the money.

Thirteen items from the last six weeks, each with my read on what it means for a business at your scale.

— Sam

Let’s get into it.

Ethercycle analyzed 24 months of raw analytics from 10 established Shopify stores, covering $274 million in revenue, 94 million sessions, and 2.2 million orders.

Year to date in 2026, AI assistants accounted for 0.10% of revenue, about $75,000 out of $76.5 million. Visitors from those assistants converted at 2.66% against a 1.91% portfolio average, and 85% of the revenue they produced came from first-time buyers. It’s a small agency-published sample using last-non-direct-click attribution, and the agency states those limits up front. (Source)

A channel that reaches strangers is the one thing your list can’t do for you, and this one is doing it at the scale of a decent month for a single store. So measure it before you move a dollar toward it.

Tag the traffic this week so it shows up in your reporting at all. Then run a monthly workflow that hands you three numbers per cohort:

If your reporting can’t separate AI-referred orders from direct today, that’s the whole job for now. It’s an afternoon with an analytics export and an agent that knows which referrers count.

Adobe Analytics reported that July 2026 AI referrals to US retail sites rose 62% year over year and converted 60% better than non-AI traffic, with a 28% higher add-to-cart rate, 14% higher engagement, and a 33% lower bounce rate. The numbers cover direct online transactions across more than a trillion US retail visits. It’s observational vendor analysis rather than a controlled incrementality test, so treat it as a cohort to track against your own margins. (Source)

Someone bouncing 33% less than your usual visitor already asked an assistant a question, got a shortlist, and clicked through to check one item on it. The page they hit decides whether they buy, so it has to carry accurate attributes and real stock status, the current price, and shipping and returns written in plain language.

Pull the ten pages your AI traffic touches most and read them the way a buyer who has already done the research would. Then compare that cohort’s gross margin against paid search before you decide what it’s worth to you.

On Shopify’s August 5 earnings call, President Harley Finkelstein said AI-referred traffic to Shopify stores and the orders coming through those channels both tripled year over year in Q2.

He also said AI search running on Shopify’s structured catalog converted at twice the rate of AI search relying on scraped product data, and put the difference down to complete specifications and attributes across a catalog of more than a billion products. These are company-reported platform metrics with no published cohort definitions or method, so treat the 2x as directional. (Source)

An assistant can only recommend the product facts it can retrieve with confidence. A missing dimension or an empty stock field keeps you off the shortlist it reads out to the buyer. Take your top twenty high-consideration SKUs and audit each one against eight fields:

Write that audit once as a Skill and re-run it every quarter against a product export. The fields you fill in this month are the ones an assistant reads next month.

Adobe reports that AI traffic to US travel sites rose 119% year over year in July 2026 and converted only 1% below non-AI traffic, against a 47% gap in July 2025. Those visits ran 21% higher on engagement, lasted 67% longer on site, and bounced 42% less than traffic from other sources. Adobe’s data is aggregated and observational, so it’s a cohort signal rather than proof of incremental bookings on any one property. (Source) (Forbes coverage)

Forty-six points of that gap closed in twelve months on the most comparison-shopped purchase in consumer ecommerce, which tells you it’s closing on your high-ticket offer too. That covers consulting, courses, and anything that goes out with a proposal attached.

The assistant is now doing the comparison step that used to happen across six of your competitors’ tabs, which means it reads your pricing and policy pages before the buyer ever does. Get your price, availability, exclusions, and cancellation terms current and machine-readable, then net out cancellations, refunds, and service load before you call the channel profitable.

Furniture retailer Povison told Digital Commerce 360 that AI referrals account for about 5% of its website traffic, and CEO Ayden Lin expects that to reach 10% by early 2027. The company is building the channel into its Cyber 5 planning and leaning on product and brand confidence instead of discounting. Both the traffic share and the forecast are executive self-reports, and the piece carries no conversion rate, incremental revenue figure, or independently verified analytics. (Source)

Five percent of traffic earns an operating plan when a concentrated sales window is four months out, and furniture is a useful case because nobody buys a sofa on impulse off a shortlist. A buyer who arrives having already done the research wants six things settled: materials, dimensions, whether it fits through the door, delivery timing, warranty, returns.

Map which assistants are sending traffic and which pages they hit, then write those six answers onto the product pages themselves. That’s what converts this traffic, and if you sell anything people consider before buying, write them before Q4 starts.

Newspack says about half of the traffic to its publisher sites comes from agentic sources that ignore the JavaScript analytics products like GA4 and Parse.ly depend on. It’s beta-testing log-file visibility with about a dozen publishers and plans an Insights product pulling audience, conversion, revenue, advertising, and app data into one dashboard. Newspack is the source of the figure, its “agentic” category blends crawlers with other machine activity, and it shows no direct revenue gain from the new analytics layer yet. (Source)

Anyone who publishes to build an audience is making decisions off a dashboard blind to half of what arrives. Three different things get blended into that one number, and each wants a different response from you:

Pull your server logs and split them by user agent, a one-evening job with a log export and an agent that knows the current bot list. Then put an email capture on the pages the assistants keep fetching, since that’s the only part of this traffic you keep.

SaaStr founder Jason Lemkin writes that his three-person team brought the number of AI agents humans had to manage down from nearly 30 toward 20, while output rose about four times.

They consolidated agents that had started to overlap, connected finance, sales, and marketing data inside a shared “10K” agent, and ran the first four finance workflows with step-by-step human review before allowing any autonomy. The output claim is self-reported and isn’t a controlled productivity measurement, though the write-up does spell out which workflows, how much managing they took, and the one bad invoice that still needed correcting. (Source)

Every agent needs a human to feed it context, and SaaStr was feeding the same context three times over to three agents that each knew part of the picture. Most people hit that wall around ten agents.

Merging them into one agent with one context file is what produced the jump, so go through your own list and find the pairs where you paste the same background into both. Consolidate the ones already producing reliable output and leave the shaky ones alone until they work.

Anything touching invoices, customers, or spend goes through a phase where it tells you what it plans to do before it does it. That’s how the bad invoice got caught.

Owner.com CEO Adam Guild rebuilt the restaurant-software company’s acquisition path around Grader, a free AI build that produces a website, upgraded imagery, video, and an SEO and CRO audit in about five minutes, replacing sales-led onboarding as the first step.

SaaStr’s Jason Lemkin reports that more than 83% of new customers now start inside an AI product and that Owner has passed $100 million ARR, alongside an internal coordination agent and a human final-publish control on ads. The account is detailed and carries a material conflict: Lemkin is a SaaStr Fund board member and an Owner investor, and nothing in it isolates what the rebuild alone caused. (Source)

Grader finishes in five minutes, which is why it works. A restaurant owner gets a real website for their own restaurant before they’ve spoken to a single salesperson, while a free trial would still expect them to go and build it.

You can run the same play at a fraction of the scale with a diagnostic, an audit, a plan, or a first draft, generated from three inputs a prospect knows off the top of their head. Build it so the output would be embarrassing if you got it wrong. Then track whether people use what it produced, because that’s the buyer who has already decided you’re worth talking to.

Okara founder Alex Morgan reports that the AI-marketing SaaS reached 300,000 monthly visitors after pivoting off its old AI-chat positioning and concentrating on X, SEO, and influencer marketing.

Influencer launches costing under $2,000 produced more than 1.8 million views, and the SEO workflow used Search Console to find pages already ranking in positions 5 through 20 and gave each one an AI-assisted update. It’s a founder self-report, and the traffic recovery followed both a positioning reset and a multi-channel push, so no single tactic gets the credit. (Source)

Positions 5 through 20 is the cheapest SEO work available to anyone with an existing site, because you’re improving pages Google already ranks. Export that Search Console slice and have an agent write one improvement hypothesis per page. There are usually only four kinds:

Ship the fixes in batches and check positions four weeks later. On the influencer side, hold it to a cost per thousand views and a landing page you can measure, because under $2,000 for 1.8 million views is a media buy and you should judge it like one.

CFO Silvia, an AI personal-finance SaaS, reported converting more than 15% of users to paid members within nine weeks of switching on paid plans. The company credits its response speed and inference economics partly to model routing, sending portfolio analysis and tax questions to different models. It’s a founder self-report with no disclosed user base, revenue total, retention cohort, or outside audit, which makes it a pricing signal rather than evidence of durable unit economics. (Source)

Fifteen percent freemium conversion is high, and the report doesn’t give you cost per active user against gross margin after inference, which is what would tell you whether it’s a business. If you’re putting a premium AI feature behind a paywall, watch four things on one cohort view: paid conversion, weekly active usage, gross margin net of model cost, and churn in the first sixty days.

Copy the model routing while you’re at it. Send classification and formatting work to a small model and keep the expensive one for judgment calls, because inference is most of your gross margin once usage grows.

Publisher Every says a new $625-per-year All Access membership tier, bundling roughly $7,000 in partner AI-tool credits, added about $9,000 in MRR in two days. Founder Dan Shipper credited AI-supported launch operations for the speed and reported that AI-segmented emails produced $25,000 in revenue. The figures are self-reported from a concentrated launch window with no retention or contribution-margin data, and the stated retail value of the credits isn’t what they’re worth to a member. (Source)

The benefit is useful on day one, which is rare for a membership tier. Somebody paying $625 gets tools they were going to buy anyway, so they can do the arithmetic themselves without taking your word for anything.

Go and secure a partner benefit that saves your member money or time in the first week, then put it in front of the segment already doing that job. Before you make it permanent, measure four things:

Two days tells you the offer reads well, so look again at month three before you build your pricing around it.

Lovable’s first designer, Nad Chishtie, writes that the AI-app-building company launched with a visual identity prototyped in 30 minutes, then replaced it with a scalable system once the product went viral.

The First Round account puts Lovable at $10 million ARR in 60 days and $100 million ARR in eight months, and independent reporting later confirmed the company crossed $400 million ARR in February 2026 with 146 employees. The essay makes no claim that the rebrand caused the revenue, so read it as a growth-operations lesson. (Source)

Fast growth compresses your decision cycles, and whatever you threw together in the first month becomes the template everyone copies by month six.

Put one person in charge of your customer-facing system now, whether that’s the brand, the site, the proposal template, or the onboarding sequence, and give them authority to replace the prototype without convening a committee. Plan that upgrade while the prototype is still working. Keep enough customer feedback running to check the new version makes acquisition or activation easier, because a rebuild that moves neither is just cost.

ViralBench runs multiple AI models on TikTok accounts twice a day, giving each model trend intelligence from Lightreel and account access through Doublespeed, with one objective: maximize views in the fitness category. Its public dashboard showed the leading Opus 5.0 account at 2.5 million cumulative views, or 78,400 views per day and 25,000 views per post, as of August 27.

It’s a live audience experiment rather than an audited business case, and it reports nothing on revenue, leads, subscriber quality, brand safety, or how a human running the same account would have done. (Source)

The result here is capacity. A model can now publish twice a day without a human touching it and pull a real audience response, which wasn’t true eighteen months ago, and so far none of it has turned into money that anyone has published. If you want to run the test yourself, bound it hard:

Run it six weeks and judge it on the metric you attached, because views are a leading indicator and the count will look impressive either way.

AI agents are sending a small volume of unusually well-prepared demand, and every case here that turned it into money did so on the strength of something that was already finished when the visitor arrived. Povison’s answer to whether the sofa fits through the door. Owner.com’s website in five minutes.

Two of these businesses took something they’d built for themselves and put a price on it. Owner.com turned an internal audit into its front door. Every priced a partner arrangement it had already negotiated.

In upcoming Cortex issues, I will outline exactly how you can quickly install quick money AI workflows or Agents that do most of all this for you.

Get more details on Cortex here

Talk soon,
Sam Woods
The Editor