Flooding the web with programmatic listicles and engineered prompts to force a chatbot recommendation mostly doesn't work. Research published this month found that AI models often hold accurate, detailed knowledge of a brand, yet discard most of it at the moment a user asks the model to actually choose — defaulting instead to a generic "category prior." Volume gets you mentioned. Structured, citable evidence is what gets you picked. So stop counting mentions and start measuring whether you survive the model's decision step.
What "sloptimization" actually means
On June 10, 2026, The Atlantic's Will Oremus published a piece coining "sloptimization": the corporate scramble to carpet the web with auto-generated listicles and planted prompts designed to nudge ChatGPT, Gemini, and Perplexity into naming your product. His Exhibit A was Shopify, which had seeded the web with 60+ listicles ranking its own platform near the top.
The logic feels intuitive. AI engines read the web, so write more of the web, and you'll show up more. Marketers have been pouring budget into exactly this for a year.
The problem is what happens after the model reads all that content.
The Decision Gap: knowing your brand ≠ choosing it
Two days after The Atlantic piece, the team at AIVO Journal published a structural post-mortem based on audits across two consumer verticals (CPG and financial services). Their finding is the part every marketer should sit with.
When they tested whether models knew the brands, the answer was yes — the LLMs held specific, accurate knowledge about ingredients, positioning, and use cases. But the moment a buyer asked the model to choose, recommend, or commit, something shifted. Mean "fact deployment" — how much of what the model knew actually showed up in the recommendation — dropped to 23%. The model routinely threw away roughly three-quarters of what it knew and fell back on a category default.
They call this the Decision Gap: the distance between what a model knows about your brand and what it deploys when it's time to pick.
Here's the kicker. The worst-performing brands in their audit weren't the least visible. They were the ones with huge first-prompt visibility and near-zero deployment at the decision stage. Content volume got them noticed. It also added noise to the model's compression pass — the step where an LLM squeezes everything it has read down to a confident, short answer. Unanchored text walls don't survive that compression. Structured evidence does.
Why volume adds noise instead of signal
Think about what an AI engine does when someone asks "what's the best X for Y." It doesn't replay every page it has crawled. It synthesizes — compressing thousands of sources into a few sentences. In that compression, repetitive, self-serving, lightly-sourced content reads as low-confidence filler. Fifty near-identical listicles saying you're the best don't add fifty units of signal; they add one weak claim repeated fifty times.
The economics underneath reinforce this. Cloudflare's analysis of AI crawler behavior found that roughly 80% of AI crawling over the trailing year was for training, not search — only around 17% was search-purpose crawling that can actually return a citation (Cloudflare, August 2025). Most of what AI bots take from your site never had a path back to you in the first place. And the imbalance is stark: Cloudflare measured Anthropic crawling about 38,000 pages for every visitor it referred back as of July 2025 (down sharply from 286,000:1 in January 2025, after Claude added web search with citations). More content fed into that machine, on its own, mostly feeds training — not recommendations.
What actually moves the needle
If volume is the wrong lever, what's the right one? The pattern across this month's research points in a consistent direction:
Lead with verifiable, specific claims. Concrete numbers, named sources, and dated facts survive compression because they read as high-confidence. A study behind the GEO research that kicked off this field found that adding statistics, citations, and direct quotations to content can lift visibility in AI answers by up to 40% (Aggarwal et al., the original "GEO: Generative Engine Optimization" paper).
Structure for retrieval and decision. Answer the buying question directly and self-containedly, so a model can lift your claim whole — not just learn about you in the abstract.
Earn third-party corroboration. A model trusts a claim that shows up in independent, credible sources far more than the same claim repeated on your own domain 60 times.
Then — and this is the part most teams skip — measure the decision stage, not the mention stage. If your dashboard still reports "share of mention" or raw citation counts, you're optimizing for first-prompt visibility, the exact metric that flattered the worst-performing brands in the AIVO audit.
This is the gap Sourceable is built to close. Sourceable tracks how your brand actually shows up across ChatGPT, Claude, Gemini, and Perplexity — not just whether you're named, but whether you're surfaced when a real buying question gets asked. That distinction between being known and being chosen is the whole game now.
How to audit your own Decision Gap this week
You don't need a research team to start. Take your five highest-intent buying queries — the "best X for Y," "X vs competitor," "is X worth it" questions your customers actually ask. Run each through ChatGPT, Gemini, Perplexity, and Claude. Note two things separately: does the model know you (mention you accurately when prompted), and does the model pick you (recommend you when asked to choose). If there's a wide gap between the two, you have a Decision Gap — and pumping out more listicles will widen it, not close it.
Then look at why the winners win in those answers. Usually it's a specific, sourced, decision-relevant claim — not sheer word count.
FAQ
What is sloptimization?
Sloptimization is the practice of mass-producing AI-targeted content — programmatic listicles, planted prompts, self-ranking comparison pages — to try to force AI chatbots to recommend a brand. The term was popularized by The Atlantic in June 2026. June 2026 research suggests the tactic often backfires by adding noise rather than signal.
Does publishing more content improve AI search visibility?
Not reliably. Models may know your brand from high-volume content but still not recommend it. AIVO Journal's June 2026 audits found fact deployment dropping to 23% at the decision stage. Specific, sourced, decision-relevant claims outperform raw volume.
What's the difference between being mentioned and being recommended by AI?
Being mentioned means the model recalls your brand when prompted. Being recommended means the model chooses you when a user asks it to decide. The gap between the two is what AIVO Journal calls the Decision Gap, and it's the metric that matters for buyers.
How do I measure AI recommendations instead of mentions?
Run your real buying queries through each major AI engine and separate "did it name me accurately" from "did it pick me." Tools like Sourceable monitor this across ChatGPT, Claude, Gemini, and Perplexity over time so you can see whether you're being chosen, not just cited.
Stop counting mentions
The old SEO instinct — more pages, more keywords, more coverage — is the wrong reflex for AI search. Models already know a lot. The fight now is over what they say when a buyer asks them to choose. Measure that, fix the evidence behind it, and you'll spend less and win more.
See how your brand performs at the decision stage across every major AI model with Sourceable.
Top comments (0)