ChatGPT quietly rewrote the rules of AI search in early March 2026. Two separate analyses published on LinkedIn - one by Lily Ray, VP of SEO and AI Search at Amsive, and one by Chris Long, co-founder at Nectiv - documented a set of structural changes to how OpenAI's model retrieves, evaluates, and cites web content. Together, they describe a system that searches more broadly, cites more selectively, and increasingly looks for authority before choosing which sources to include in a response.

The first data point came from Lily Ray, who posted her findings on March 5, 2026, citing visualizations by Olivier de Segonzac at Resoneo. According to Ray, ChatGPT 5.3 - launched on March 3, 2026 - is linking out to websites substantially less than previous model versions. The chart shared in her post shows a visible drop in average unique domains per response after the 5.3 model release, with the "cliff" in the graph aligned with the model launch date. "So your tiny sliver of traffic from ChatGPT is about to get tinier," Ray wrote. She noted that average unique domains per response fell from approximately 19.8 before the update to around 15.9 after it - a decline of roughly 20% - while the number of responses analyzed held in the hundreds.

That drop matters because ChatGPT's referral traffic to publishers has already been under pressure for months. A 52% decline in referral traffic was recorded in July 2025 when OpenAI adjusted how its retrieval-augmented generation system weighted sources. The new 5.3 data suggests that pattern has continued into 2026.

What ChatGPT 5.3 is actually doing differently

The second piece of analysis came from Chris Long, who documented a parallel shift: ChatGPT is now issuing substantially more search queries per response. According to Long, earlier research at Nectiv found ChatGPT averaging more than 2 queries per response, with a maximum of 3. In the 5.3 model, that number has jumped significantly - the system is now "oftentimes searching 10+ different fan-outs," as Long put it. The term "fan-out" refers to the branching of a single user query into multiple sub-queries that the model runs independently to gather information before composing its response.

More queries but fewer outbound links. The combination points to a retrieval architecture that has become more rigorous about which sources qualify for inclusion, even as it consults more of them during the research process.

Long documented several specific behavioral patterns he observed in GPT-5.4 Thinking - a variant of the model available to paid users, with the standard version for most users being 5.3 Instant. The distinction is important: Long noted that the default model for the roughly 95% of users who access ChatGPT through the free tier is 5.3 Instant, not the Thinking variant. The behaviors he documented may therefore not apply uniformly across the entire user base.

Site: operator and direct brand lookups

One of Long's most concrete observations was that GPT-5.4 is using the "site:" operator more frequently when researching queries. In standard web search, the "site:" operator restricts results to a specific domain. According to Long, ChatGPT is now using this to find information from the brands themselves, rather than relying on third-party coverage. A search about a specific company might generate a fan-out query like "site:companyname.com product specifications" rather than general web search.

This represents a shift from information discovery to information verification - and it changes which pages get retrieved. A brand's own product pages, documentation, and official content become more directly relevant to AI citation, independent of how those pages rank in organic search. The observation aligns with research showing that ChatGPT citations often reference content ranking in traditional organic search positions 21 or lower approximately 90% of the time- a pattern that already contradicted conventional SEO assumptions.

Authority signals: accreditation, awards, third-party validation

Long also documented ChatGPT searching for explicit authority signals that were not mentioned in the original user query. For a query about "best nursing programs," the model independently searched for NCLEX pass rates and CCNE accreditation - two external validation signals not asked for by the user. For a query about "best SEO agencies," the model searched for Search Engine Land award winners and checked profiles on Clutch and G2, which Long described as trusted data providers and aggregators in the SEO industry.

This behavior - generating authority-checking sub-queries unprompted - suggests the model has incorporated something resembling a trust verification step into its retrieval pipeline, at least for category queries where quality signals are available. The practical implication is that brands without documented accreditation, award recognition, or presence on established aggregator platforms face a structural disadvantage in these types of queries, regardless of the quality of their website content.

Long's assessment of the results was direct: "The answers are BETTER." He noted that "best SEO agency" queries previously returned lists of agencies he had never heard of, whereas the new model surfaces firms like Seer Interactive, Amsive, and iPullRank - which he described as actual leaders in the space. The improvement, in his characterization, came specifically from the model prioritizing verifiable authority over content volume or keyword density.

Why traffic and citation are decoupling

Ray's commentary raised a dimension of the 5.3 update that goes beyond citation counts. She acknowledged the common counterargument - that AI visibility is about brand mentions, not just outbound links - but maintained that the decline in linking behavior still carries weight for "traffic-obsessed CMOs" or any organization measuring success through "citations from ChatGPT" as a performance metric. Her framing was pointed: "Brands can be recommended in ChatGPT but not have much to show for it in terms of referral traffic, and this is true now more than ever."

The gap between brand recommendation and referral traffic is not new. Research from November 2025 showed that AI referral traffic grows quickly but converts at rates far above traditional channels - with visitors from large language models converting to sign-ups at 1.66% compared to 0.15% from search. That conversion premium becomes harder to capture when the model mentions a brand but does not link to its site. The 5.3 update appears to extend this dynamic further: brands may appear in AI responses without generating a single referral visit.

The technical architecture behind the changes

The WordLift blog published a detailed technical analysis on March 6, 2026 - written by Andrea Volpini and cited by Lily Ray in her LinkedIn post comments - that provides structural context for understanding how these retrieval changes work under the hood. According to Volpini, Perplexity's research on its pplx-embed model makes one architectural fact explicit: "Citations start with retrieval." In Perplexity's pipeline, embeddings are used at the first stage of candidate selection for web-scale retrieval. Documents that are not retrieved into the candidate set cannot be cited.

The article identifies three distinct technical paradigms across the three major AI search systems. Perplexity uses bidirectional context through diffusion-based continued pretraining, along with support for "late chunking" - an approach that embeds passages with awareness of the full document rather than treating each segment in isolation. According to the analysis, this means mixed-intent pages or loosely connected sections risk diluting the contextual signal injected into each chunk.

Google's approach, according to the article, involves passage-level ranking via BERT and successor models. A single highly relevant paragraph buried deep in a disorganized page can independently trigger retrieval. This creates an asymmetry: the system is looking for text that behaves structurally and semantically like an answer. Content that too closely mirrors query phrasing risks behaving more like a query than a document - which reduces its competitiveness in retrieval.

OpenAI's text-embedding models use a technique called Matryoshka Representation Learning (MRL). This allows an embedding vector of up to 3,072 dimensions to be truncated to as few as 256 dimensions without losing core conceptual meaning. The most broadly significant semantic information is mathematically front-loaded into the earliest dimensions of the vector. For content producers, this has a structural consequence: sections that bury the core thesis in a fourth paragraph may have that thesis truncated out of the semantic vector during the initial fast retrieval sweep, before the model even reaches the reranking stage.

What gets cited and what does not

According to Volpini's analysis, content that achieves citation tends to share identifiable structural characteristics: semantically explicit passages, self-contained sections, entity-grounded writing, evidence-backed claims, and topical coherence at the document level. Each of these qualities maps to a measurable signal in the retrieval systems described.

The article draws on additional research by Kevin Indig, cited as a large-scale analysis of how ChatGPT cites sources. According to that research, a large share of citations come from the early portion of a page, and citations correlate with passages that are definition-heavy, entity-dense, and factual. The WordLift analysis concludes: "Citation behavior is downstream of retrieval competitiveness." If a passage is easy for machines to understand and trust, it is more likely to be retrieved, selected, and cited.

This framing has direct implications for how the 5.3 changes should be interpreted. The reduction in outbound links is not simply a product decision by OpenAI - it appears to be a byproduct of tighter retrieval standards. Sources that pass a higher trust and clarity threshold get included; those that do not are dropped from the output even if they were consulted during the research process.

Authority abuse and the spam problem

One recurring theme in the comment threads under both LinkedIn posts was concern about manipulation. Multiple commenters noted that agencies are already testing how to manufacture the authority signals that ChatGPT now checks - engineering Clutch profiles, creating accreditation-style pages, and reverse-engineering the award references the model finds. One commenter observed the pattern directly: "Noticed the same and also noticing some agencies already abusing it."

Research published in February 2026 documented how companies pay to influence what ChatGPT recommends, with some practitioners placing strategic content across multiple domains to shape AI responses. The WordLift analysis addresses this directly, noting that Google's March 2024 spam policies focused on scaled content abuse, expired domain abuse, and site reputation abuse - practices it describes as "spam patterns wearing new clothes." Modern retrieval stacks, the analysis argues, evaluate signals that correlate with usefulness - semantic specificity, entity clarity, informational completeness, factual grounding, structural coherence. Content that is thin or overly templated produces weaker semantic signatures in vector space.

What this means for the marketing community

The combination of more search fan-outs and fewer outbound links creates a measurably different competitive environment for brands trying to maintain visibility in AI search. The model is doing more research per query - not less - but is filtering more aggressively at the output stage. A brand that lacks accreditation pages, award documentation, or profiles on recognized aggregators faces reduced citation probability on category queries, regardless of organic search rankings.

LinkedIn's position in AI citation rankings - ranking second overall across AI platforms, cited in 11% of AI responses on average - points to one practical consequence: the platforms where authority signals accumulate publicly, such as LinkedIn profiles and company pages, G2 reviews, and accreditation registries, are increasingly the sources AI systems consult when checking credibility. The Semrush research on LinkedIn citations, covering 325,000 unique prompts submitted between January and February 2026, found that 54-64% of cited LinkedIn posts focus on sharing knowledge or practical information, not promotion.

For marketing teams measuring AI search performance, Ray pointed to brand-name search demand as a more reliable indicator than citation counts or referral traffic. Search impressions for branded queries in tools like Google Search Console - where Google can track impressions - offer a more stable proxy for AI-driven interest than raw ChatGPT referral numbers, which have proven volatile across multiple model updates. The caveat she added was pointed: "Well, assuming Google can get its act together with tracking impressions."

The broader shift from keyword-based retrieval toward credibility-first selection has been tracked across multiple data points on PPC Land, including the finding that citation sets change 40-60% within a single month and that blocking AI crawlers does not reliably reduce citation exposure. The instability of citation patterns makes long-term strategy difficult to build around specific citation outcomes - which is precisely why Ray and others in the discussion thread pointed toward structural authority building rather than citation optimization as the more durable approach.

Timeline

Summary

Who: Lily Ray (VP of SEO and AI Search at Amsive, founder of Algorythmic), Chris Long (co-founder at Nectiv), Olivier de Segonzac (Resoneo), and Andrea Volpini (WordLift) published the analyses. OpenAI is the company behind the ChatGPT model changes.

What: ChatGPT 5.3, launched March 3, 2026, links to fewer websites per response - with average unique domains declining from approximately 19.8 to 15.9 per response - while simultaneously searching 10 or more fan-out queries per prompt, up from a previous average of 2. The model uses the "site:" operator to look up brand-owned content directly and checks authority signals such as accreditation credentials, industry awards, and aggregator profiles even when not asked to.

When: ChatGPT 5.3 launched on March 3, 2026. The Resoneo data and Lily Ray's LinkedIn post were published on March 5, 2026. The WordLift technical analysis was published on March 6, 2026. Chris Long's LinkedIn post was also published within approximately the same two-week window.

Where: The behavioral changes affect ChatGPT's web retrieval across all query types, with the most documented effects on category queries such as "best SEO agency" and "best nursing programs." The default model for free-tier users is 5.3 Instant. The Thinking variant documented by Long is available to paid users. Effects are visible in referral traffic data tracked by Resoneo and analyzed against Bing Webmaster Tools, Google Search Console, and third-party analytics platforms.

Why: The changes reflect OpenAI's effort to improve response quality by checking authority and credibility signals more rigorously before including sources in outputs. The WordLift analysis provides the underlying technical explanation: embedding architectures from Perplexity, Google, and OpenAI all favor passages that are semantically explicit, entity-grounded, and factually verifiable. Content without those qualities scores poorly in retrieval candidate selection - and therefore does not reach the synthesis stage where citation decisions are made.

Share this article
The link has been copied!