Surfaced - AI Search Intelligence - October 1st, 2026
Attribution needs trust, AI platforms have different incentives, and the technical work worth doing this week
By Andrew Riker, Founder, Surfaced
Took a trip to Yellowstone last week, detoxing from AI is real.
Bison, bald eagles, bears, trout, moose, all on top of a supervolcano. Excellent trip with family and friends, a reminder that experiences cannot be replaced through the internet.
Anyway, you’re here for business. Kevin Indig’s weekly newsletter was a banger this week: how useful is attribution for AI search?
Then Rand Fishkin showed up in my feed, taking a hammer to “How did you hear about us?” surveys. Rand going against the grain? Shocking 😅
I agree with the direction. Incrementality tests and patterns across several measures can tell us something useful when the click trail falls apart. But there’s an organizational step that gets skipped: you need people to trust the work before they’ll fund something they can’t trace neatly to a conversion.
“Trust me bro” isn’t enough.
This issue connects that problem to the platforms themselves, what they’re building, how they make money, and a few things your team can actually improve this week.
I’ve also been talking through a few of these challenges on LinkedIn. Catch the videos and add your take:
Also, I wrote an article on Taste, Artistry, and Judgment. Humans still matter.
Onward!
Same category but verrrry different incentives
WIRED reported that Perplexity was stepping away from advertising, with executives citing the risk to trust in its answers and a renewed focus on subscriptions. Big jump from what they originally were positioned.
A few weeks ago on August 31, OpenAI reported that ChatGPT Ads had reached a $1 billion annualized revenue run rate.
These are opposing choices about advertising, though OpenAI also sells subscriptions and enterprise products. OpenAI says ads are labeled, separate from answers, and do not influence those answers. Suure.
My take: your AI search plan needs to name the surface. An ad-supported product has reasons to develop advertiser tools and commercial placements. A subscription-led product has reasons to keep paying users satisfied enough to renew.
This goes the same for social platforms, which Reddit and YouTube are pushing to enable AI-enabled search directly in an attempt to bypass your usual “reddit” or YouTube appendix on a Google search.
The infrastructure beneath the announcements
Tomek Rudzki’s September 4 research at Peec AI describes OpenAI’s own family of search indexes, alongside outside providers. He also found experiments comparing its index with third-party results, especially in shopping.
Source: Peec AI. Screenshot of a separate shopping retrieval experiment discussed in the research.
The question is how much retrieval shifts toward OpenAI’s own infrastructure and away from SERPs they’ve crawled to get data in the first place. If that share grows, especially for commercial queries, the citation patterns we’ve spent the last 18 months mapping could change substantially and your world might turn upside down a bit.
OpenAI taking punches at Google’s incumbency, love it.
Google has invested decades into search, and they will continue to get bites taken out of them from various LLMs. OpenAI knows it’s a data war, and they are acting accordingly.
Nothing to do at the moment but monitor market share and global usage.
Rand is poking the attribution bear again
In this LinkedIn post, Rand Fishkin argues that checkout discovery surveys can’t reconstruct how someone bought. People forget, favor recent or familiar channels, and compress a lot of influences into one answer. A survey also won’t reveal an opportunity in a channel you’ve never invested in, so it’s fundamentally borked.
I appreciate how often Rand pushes against the comfortable industry answer. And I agree that incremental lift deserves more attention and there are various folks coming to the same conclusion.
Where I want to spend more time is the handoff to leadership. You can be right about attribution’s limitations and still lose credibility if the replacement sounds like “the numbers went up around the same time.”
Correlation helps us choose what to investigate, and a well-designed experiment gives us a stronger basis for a decision.
Before you ask for a different measurement approach, get agreement on what the test can establish, what it can’t, and what result would change the budget.
Attribution is under attack
“The usefulness of attribution for AI Search” arrived this Monday, featuring George Bonaci, VP Growth at Ramp.
Kevin’s argument is click attribution captures a shrinking slice of demand, and I agree.
1) Keep the observable data, then combine exposure, customer behavior and business outcomes to judge progress.
2) Use incrementality tests where you can construct a credible comparison.
3) Several signals moving together strengthen the evidence, but don’t establish causality.
That’s a lot of words, but ultimately making short tests is the way to go. Lagging indicators are broken due to cookies, buying journeys, attribution models, mobile vs. desktop, and 100 other reasons.
Unlike Rand, Kevin includes self-reported discovery among the imperfect signals you can combine.
I’d keep customer conversations for context and use experiments to support budget decisions, absolutely!
Graphite put numbers behind “this sounds like AI”
Graphite’s September AI Tells research compares 10,000 human articles with 90,000 AI-generated articles across nine models. The writing habits differ by model family and change between versions. Very neat study.
Source: Graphite, AI Tells, September 2026. Em dash use varies by model.
Even the familiar em dash test gets messy 🙃 . In the study, some current models use it much less often than humans. The study uses one fixed prompt and older human articles, so use it to sharpen your editing.
My take: removing a few phrases won’t give your writing a point of view. Put in the specific observation, the tradeoff you’d make, the thing you disagree with. I think us hoomans are getting better at finding AI-generated stuff, and there’s a strong gag reflex to it. As for robots, they care less if it’s of value.
Chris Long at Nectiv made a banger
Chris Long’s September 21 AEO page guide pulls crawl access, clear takeaways, useful structure, fresh facts, customer questions and related subquestions into one visual. Great resource to put next to a page your team is already working on. I could recreate it, but nope.
I’d use it as a review checklist, especially if you need to explain what to optimize a page for - however, it doesn’t capture the domain level and what a site needs, just tactical recs.
Can a buyer get the answer quickly, verify it, and find the next detail they need?
I’ve been using Instinct and Muse for Personal Assistant’s lately, and it’s amazing how much it can do for me without me ever visiting websites. With Cloudflare heralding the invasion of agentic visits and consumers getting their hands on products to make their everyday worthwhile… I ask:
Can an AI agent actually use your website?
Give an AI browser agent a real customer task, like
“Find the right product”
“Compare two plans”
“Complete a demo-request form in a test environment”
Can it finish, and can it tell that it finished?
That’s a useful way to think about agent readiness. Getting your content read is one part of it. Helping an agent take the next step on a customer’s behalf brings your forms, navigation and interactive controls into the picture.
Layout stability belongs in that conversation. Anthropic’s description of computer use explains how an agent interprets screenshots and chooses click coordinates. If a button moves between the screenshot and the click, there’s a plausible risk of a wrong action.
That’s where SEO/Core Web Vitals come into play!
There’s a concrete engineering connection, too. Playwright checks that a click target is visible, stable, enabled and unobstructed. If those checks don’t pass before the timeout, RIP.
Agents using that approach can handle some movement, but persistent movement or an overlay can still stop the task.
The assignment this week: test three customer journeys in with your favorite agent or PA.
Choose tasks with a clear finish line. Pick journeys that matter commercially, such as finding a suitable product, comparing plans and requesting a demo. Write down what a correct result looks like. Use a test environment for submissions or purchases.
Run each task several times. Keep the instructions, agent version and starting conditions consistent. Record completion rate, wrong actions, retries, time to finish and where a person had to intervene. Save the trace or recording so engineering can reproduce the failure.
Watch the whole interaction. Look for moving buttons, late-loading banners, overlays, unclear field labels, awkward date pickers and missing confirmation messages. Check after scrolling, opening menus and triggering validation. A page-load score won’t capture every problem in a customer journey.
Fix patterns across templates. Reserve space for images and embeds, stabilize key controls, use clearly labeled buttons and form fields, and make loading, errors and success states explicit. Group failures by shared component so one fix can improve several journeys.
Repeat the same tasks after the fix. Compare completion and intervention rates with the baseline. Expand to other templates and agents once the first journeys work reliably. Keep the agent and test conditions in the report so you know what the result applies to.
Keep Core Web Vitals in the review, especially unexpected layout shifts. But a good CLS score alone doesn’t establish agent readiness, and improving it doesn’t come with a proven AI citation lift.
❝Should we buy an AI visibility platform, or build our own?
(Usually) Buy the monitoring. Build the strategy and the diagnosis.
Monitoring is turning into a commodity, and pricing is heading down and your team should not be maintaining prompt-tracking scripts.
The strategy a platform, LLM, or any other software outside your company gates couldn’t possibly understand.
The diagnosis is specific to your business, and no vendor can package your judgment.
What people skip is what building costs you later. Every script, app, dashboard, and internal tool is something somebody now owns.
Ten of those and you’re running a small software company and likely under crumbling tech debt. Technical debt is very real, maybe you’re like me and juggling a dozen apps you dont even know how to manage.
Which makes the deciding factor organizational rather than technical. Three things matter:
Can your people actually use these tools, or is it one person with a prompt they won’t share?
Can they make small things? A one-page app, an artifact, a script that answers a single question and then gets thrown away.
Is your company context written down anywhere an AI can reach it? Positioning, products, competitors, how you describe yourselves. Without that you get generic output no matter how good the model is.
I built a site audit in an afternoon for pennies of compute. It found a company where 74 pages out of 3,730 described what the business actually sells.
No platform surfaced that, and it used embeddings to assess a page's relevance beyond keywords.
It came from knowing which question to ask, something vets are going to always be stronger at.
It sure is fun building, but the management is giving me a huge headache 😇
Until next time,
Andrew Riker
This issue was first sent to subscribers of the Surfaced newsletter on October 1, 2026.