The Desk

August 11, 2026

GPT-5.6 Made Every AI-Visibility Friction Twice as Costly

David Konitzny at Peec AI measured GPT-5.6's retrieval shift... every Cloudflare block, protocol gate, and stale page between your content and AI just got twice as costly.

Two pieces on The Desk right now cover the same thing from different angles (I have really been looking forward to saying that... "on The Desk").

One is about how Cloudflare's managed robots.txt setting can silently block the AI crawlers your dashboard says are allowed. The other is about how the "merchant consent" framing in agentic commerce protocols recreates the friction the open web spent thirty years removing.

Then David Konitzny, a GEO researcher at Peec AI, published data that sharpens the point of both.

His research, based on network-traffic analysis of ChatGPT's search architecture, shows that GPT-5.6 fundamentally changed how the model retrieves web content. The changes are large. Roughly 2x on the metrics that matter most for AI visibility.

OpenAI has not publicly discussed any of these changes. They rolled out capability benchmarks and pricing updates.

What OpenAI announced with GPT-5.6

Standard model-release marketing. GPT-5.6 launched July 9, 2026, as a family of three models: Sol (flagship), Terra (balanced), and Luna (cost-efficient). On August 6, 2026, Luna became the default for Free and Go users.

The announcement emphasized benchmark gains in coding, biology, chemistry, and cybersecurity. It talked about pricing (Luna dropped 80% on July 30). It talked about long-context Fast mode. It talked about capabilities.

It did not talk about how the model retrieves content.

What Konitzny measured

Per Konitzny's analysis of ChatGPT's network traffic before and after the GPT-5.6 default rollout, five things shifted.

Single-iteration retrievals dropped from 94% to 43.5% of prompts. In GPT-5.5, most prompts got one round of search and one set of sources. In GPT-5.6, more than half of prompts trigger multiple rounds of search, with the model refining its query based on what the first round returned. Two-iteration prompts rose from 5% to 33.5%. Three-iteration prompts rose from 0.1% to 7.4%.

Average sources per prompt roughly doubled, from 12 to 24. Every prompt now pulls in twice as many URLs as reference material.

The site: operator jumped from 0.3% to 23% of fan-out queries. GPT-5.6 is now actively restricting significant portions of its search to specific domains, which means brand websites, not just aggregators, are being targeted directly.

Product pages became the dominant source type, up 119% in daily retrievals. Listicles dropped from 15.47% to 8.19% of sources. The model is pulling from first-party product information over roundup articles.

This one has massive implications for agentic commerce. Product Detail Page (PDP) optimization gets more important with every model improvement, and more important with every shopper who uses ChatGPT or Claude instead of Google to research a purchase. When a shopper asks an assistant "what's the best winter jacket under $300," GPT-5.6 is now more likely to pull from actual brand PDPs than from a Buzzfeed roundup. Which means every merchant's PDP is now their front door to AI-assisted shopping, and the merchant whose PDP loads slowest, hides price in JavaScript, or has no schema markup is the one who gets skipped over.

Recency signals in searches climbed from 10-15% to 28.6%. Roughly triple. Year-based queries are now dominant, meaning GPT-5.6 routinely appends the current year to its own search terms ("best CRM 2026," "agentic commerce protocols 2026") to filter for current information over stale content.

What this means practically: a page from March 2024 with no updates is now competing against a page from last week for the same query and losing more often. If your dateModified is stale, or your content doesn't reference the current year anywhere, GPT-5.6 has an easier time picking someone else. Every page you want retrieved needs to look current to a machine that is now three times more likely to filter for currency.

The full data table is in his LinkedIn Pulse post.

How he measured it

None of these numbers came from OpenAI. Konitzny reverse-engineered them from network traffic.

Peec AI's methodology, which Konitzny helped develop, captures the WebSocket stream that ChatGPT's UI displays as the "Researching..." panel. They open browser DevTools before launching a research session, filter to WebSocket traffic, and record raw frames for offline analysis. They then verify their captures against independent page fetches (the agent's log of what it "read" from a page matches an unauthenticated crawl of the same URL, word for word).

Their publication rule is that a finding has to repeat across ten or more separate accounts and multiple prompt types before they treat it as real. This is technical SEO reverse-engineering by someone who has spent years doing it professionally, most recently as Head of Organic Search at a European fintech. It is not hot takes.

Also worth noting: Konitzny's title at Peec AI is GEO Researcher... which, for what it is worth, should help settle the GEO naming debate. ;P

It is also not officially confirmed by OpenAI. Any specific number in this piece attributed to Konitzny is a measurement, not an announcement.

The convergence

Recently I have been focusing my research on understanding readiness in agentic commerce in practical ways... and three recent topics have all been converging.

The first is Cloudflare. A managed robots.txt setting that most operators have never touched (and often did not know was on) can silently block GPTBot, ClaudeBot, and CCBot at the file layer, while the same AI Crawl Control dashboard shows every one of those crawlers as allowed. The plumbing looks correct. The dashboard confirms the plumbing looks correct. The crawlers actually skip you.

The second is the merchant consent model in agentic commerce protocols. UCP, ACP, Shopify Agentic Storefronts all require merchants to activate, enroll, upload data, install apps, approve agents. Commerce on the open web has never worked that way. Any customer with a valid payment and a checkout that accepts it can complete a purchase, and merchants gate the transaction (fraud check, inventory, MAP pricing) rather than the demand channel.

The third is Konitzny's data. GPT-5.6 doubled its source retrieval, tripled its use of site: targeting, and made product pages the dominant source category.

The three point at the same problem.

AI systems now rely on direct content from first-party sources. They are targeting brand websites specifically. They are pulling twice as many URLs per prompt... the friction between the model and the content is where the value gets destroyed. That friction takes many forms: Cloudflare blocks, protocol enrollment gates, stale product pages without a recency signal.

Every friction now costs roughly 2x what it did a week ago, because the model is retrieving roughly 2x as much.

What this means for agentic commerce

Let's apply this to a practical scenario with ChatGPT...

A shopper opens ChatGPT and asks a product question. Under GPT-5.5, the model ran roughly one search, pulled in twelve sources, mostly from listicles and roundup articles, cited a few. Under GPT-5.6, the model runs multiple search rounds, pulls in twenty-four sources, targets specific brand domains with site: restrictions, prefers product pages over listicles, and prioritizes recent content.

If your brand's product page is indexed, retrievable, and current, the model finds it. Without you enrolling in anything. Without any protocol adopter needing to approve you. Without any intermediary gating whether the model can quote your inventory...

If your product page is blocked, buried, or six months old with no recency signal, the model finds someone else.

What to do this week

If you have a Cloudflare-fronted site with a stake in AI visibility, the shortest path to acting on Konitzny's data is three steps.

Check your robots.txt. Run curl -sL https://YOURSITE.com/robots.txt | head -20. If you see BEGIN Cloudflare Managed content, disable it in your AI Crawl Control settings. The full walkthrough is in the Cloudflare Content Signals piece.

Refresh your dateModified. GPT-5.6's recency preference roughly tripled. Any page you want retrieved should reflect current information and current dates in its structured data. The five-line SEO fix is bumping <meta property="article:modified_time"> and dateModified in the JSON-LD schema on every page you want the model to treat as current.

Audit your product pages. Listicles are now the wrong bet. While I have not had a career in SEO, I am surprised it has taken this long for LLMs to start disliking listicles... they have always seemed like a cheap way to game SEO. Odd it took this long. Regardless: if your public brand pages are the primary place customers can read your product information, they need to load fast, present price and availability in HTML (not JavaScript-rendered), and expose schema.org Product markup. If you have not looked at your product pages through an agent's eyes recently, do it this week.

None of this requires protocol adoption. None of it requires enrollment... it requires making sure that when GPT-5.6 comes looking, twice as often, in three rounds, restricted to your domain, it can actually read what you have.

The retrieval shift is a good thing

For anyone who was frustrated by the drift toward walled gardens and platform intermediaries in agentic commerce, Konitzny's data is a signal that the direction is still salvageable. The models want direct sources. They want first-party product information. They want current inventory.

They want the open web to work. And they are not retrieving from the protocols.

They are also willing to look harder for it, twice as many URLs per prompt, in three rounds, before giving up.

The infrastructure between AI and your content is what determines whether they find you... and that infrastructure is under your control. :)

Frequently Asked Questions

What did OpenAI announce with GPT-5.6?

GPT-5.6 launched July 9, 2026 as a family of three models: Sol, Terra, and Luna. Luna became the default for Free and Go users on August 6. OpenAI's announcement focused on capability benchmarks and pricing, not on how the model retrieves web content.

What did David Konitzny measure about GPT-5.6's retrieval behavior?

Five shifts: single-iteration retrievals dropped from 94% to 43.5%, sources per prompt roughly doubled from 12 to 24, the site: operator jumped from 0.3% to 23% of fan-out queries, product pages became the dominant source type (up 119% daily) while listicles dropped from 15.47% to 8.19%, and recency signals in searches roughly tripled from 10-15% to 28.6%. Full data in his LinkedIn Pulse post.

How was GPT-5.6's retrieval behavior measured?

Peec AI captures the WebSocket stream ChatGPT's UI shows during research sessions, verifies findings against independent page fetches, and only publishes what repeats across ten or more accounts. It is reverse-engineered from network traffic, not official OpenAI telemetry.

Has OpenAI officially confirmed the retrieval changes in GPT-5.6?

No. OpenAI's public communications focus on capabilities, benchmarks, and pricing. The specific retrieval changes (multi-iteration search, source volume doubling, site: operator use, product page dominance, recency preferences) are Konitzny's measurements, not OpenAI announcements.

What does the GPT-5.6 retrieval shift mean for AI visibility and agentic commerce?

GPT-5.6 targets brand websites via the site: operator, prefers product pages over listicles, prioritizes recent content, and pulls roughly twice as many URLs per prompt as GPT-5.5. For anyone in agentic commerce, every friction between the model and your content (Cloudflare blocks, protocol enrollment gates, stale pages) now costs roughly 2x what it did before.

What should merchants and content owners do to prepare for GPT-5.6's retrieval?

Three steps. Check your robots.txt for the Cloudflare Managed content block and disable if present. Refresh your dateModified fields since GPT-5.6's recency preference roughly tripled. Audit your product pages: server-rendered HTML for price and availability, plus schema.org/Product markup. No protocol adoption required.

Source