What ChatGPT Reads Before Deciding Not to Cite You

Most optimization advice for ChatGPT citation visibility assumes the problem is retrieval. Get into the right directories. Use schema markup. Build backlinks from sources ChatGPT trusts. Get those things right, and you'll show up.

That assumption is close to correct, but it misses a second gate that runs after retrieval -- one that doesn't look at your content at all.

The Stage You Already Know About

ChatGPT doesn't retrieve pages directly from a user's query. First, it generates a set of web search sub-queries from the original prompt. Then it retrieves pages matching those sub-queries. The result is a pool of fetched URLs, typically around 16 to 17 per prompt. If you're not in that pool, you're invisible.

This is the stage most AEO advice addresses. Directory presence helps your pages enter the retrieval pool. Schema markup helps them rank in the retrieval queries. The 5W PR HVAC/Plumbing AI Visibility Index (65+ prompts across ChatGPT, Claude, Perplexity, and Google AI Overviews, January-March 2026) found 87% of independent contractors have zero citation share in their own market. In most cases, they're failing at this first stage -- no directory infrastructure for ChatGPT's retrieval queries to find them.

But retrieval alone isn't enough. Here's what happens after pages enter the pool.

The Filter ChatGPT Runs Before Opening Your Page

In our 2026-10-16 review of the Ahrefs 1.4 million ChatGPT prompts study (Ahrefs primary research, February 2026 data, published April 2026), we documented a specific mechanism that determines which retrieved URLs actually appear in ChatGPT's final response.

ChatGPT retrieves an average of 16.5 URLs per prompt. Of those, roughly 50% get cited and 50% do not. The number of URLs retrieved per prompt is nearly identical whether a page is later cited (16.57) or not (16.58). Getting retrieved frequently does not predict getting cited.

The filtering decision is made primarily using metadata available from the retrieval result -- specifically the page title, URL slug, and snippet -- before full page content is opened. Two signals dominate the selection:

**Title-to-sub-query semantic alignment** is the primary citation selection driver. If the page title directly answers the specific sub-query that retrieved it, it passes the filter. If the title is generic relative to that sub-query, it competes against 16 other pages in the pool and loses most of those competitions.

**URL slug readability** is the second measurable signal. The Ahrefs data shows an 8.67 percentage point gap: pages with natural language URL slugs (like `/emergency-hvac-service-dallas/`) cite at 89.78%, while pages with opaque URL patterns (like `/service?id=4721`) cite at 81.11%. The mechanism is consistent with how the filter works -- readable slugs signal what a page contains without requiring content access, and ChatGPT is making this evaluation at the metadata stage.

Why Title Precision Matters More Than Keywords

The title alignment finding makes more sense once you understand the underlying retrieval architecture. In our 2026-06-26 analysis of the fan-out mechanism (from Nick Lafferty's research on ChatGPT query architecture), ChatGPT generates 91% unique sub-queries from a single user prompt. The retrieval pool is drawn from this fan-out of sub-queries, not from the primary query.

This creates a specific mismatch most businesses don't account for. A user asks: "Who should I call for a broken furnace in December?" ChatGPT doesn't search that query. It generates sub-queries like "emergency furnace repair 24-hour service," "licensed HVAC technician winter heating calls," "furnace breakdown same-day service near me." Each sub-query retrieves its own set of pages.

A page titled "HVAC Services and Repairs" gets retrieved if it appears in a broad directory listing that ranked for one of those sub-queries. At the metadata filtering stage, the title competes against pages titled "Emergency Furnace Repair 24-Hour Service" and "Same-Day Heating System Repair." The generic title loses.

This is not a content quality problem. It's a title specificity problem. The fix is not rewriting the page body -- it's ensuring the page title directly reflects the specific angle of the sub-queries likely to retrieve it.

Reddit Gets Retrieved, Then Dropped

Reddit illustrates this filter mechanism clearly.

ChatGPT retrieves a significant volume of Reddit content for context, but the Ahrefs data shows a 1.93% citation rate for Reddit-sourced URLs -- the lowest channel rate in the study. 67.8% of all non-cited URLs in the 1.4M prompt dataset came from the Reddit source.

The explanation is not that ChatGPT's editorial layer actively suppresses Reddit. Reddit URL slugs are frequently opaque (reddit.com/r/HVAC/comments/x8v42j/), and Reddit post titles are conversational rather than semantically precise to the sub-queries that retrieved them. Reddit provides retrieval volume but fails the metadata alignment stage consistently.

This has direct implications for any AEO strategy that emphasizes Reddit presence as a ChatGPT lever. The October 2025 Reddit citation collapse was driven by a search parameter change that reduced Reddit's retrieval presence. But even before that, the Ahrefs data from February 2026 shows Reddit was already being retrieved heavily and cited at under 2%. Getting retrieved on Reddit provides almost no citation benefit for ChatGPT.

How Gemini Filters Differently

The post-retrieval filter is not universal across platforms. Our 2026-09-23 analysis of the Steady Demand AI Citation Ledger (14,472 local citations across 50 US metros and 10 local service categories, data from July 2026) showed that 59.9% of Gemini's local citations come from the business's own website.

For Gemini, website authority and semantic structure dominate citation selection. Title alignment and URL readability still matter, but the primary lever is the quality and structure of the brand site itself. A business with a strong, well-structured website earns Gemini citations at a much higher rate than the same business with a thin site but strong directory presence.

ChatGPT is the inverse. Third-party directories account for 49% of ChatGPT's citations (Yext 2026, 17.2M citations). ChatGPT's filter evaluates directories, comparison articles, and third-party sources far more than brand sites. If your brand site has specific, readable page titles and clean URL slugs, those help at the selection stage -- but the majority of ChatGPT citations come from third-party sources where you have no control over URL structure.

The implication for sequencing: for ChatGPT, title and URL work only after directory presence work. Getting those metadata signals right doesn't compensate for not being in the retrieval pool.

What Changes in Your Fix Plan

Three specific changes follow from this data.

**Title precision over keyword density.** Each service page should carry the specific vocabulary of the sub-query most likely to retrieve it, not a generic keyword grouping. "HVAC Services in Toronto" competes badly in a retrieval pool alongside pages titled "Emergency HVAC Repair Toronto" and "24-Hour HVAC Technician Toronto Winters." If the fan-out generates sub-queries about emergency availability or winter-specific services, only the more specific title consistently passes the filter.

**URL slug cleanup as a Phase 1 infrastructure task.** Any page with an opaque URL pattern -- query strings, numeric IDs, auto-generated admin paths -- has a measurable citation disadvantage according to the Ahrefs data. This is a technical change that requires no content work and produces an 8.67pp lift at the citation selection stage. It should sit in Phase 1 infrastructure work alongside schema cleanup and NAP normalization, not in a later optimization phase.

**Diagnosis at both stages.** A page that is being retrieved but not cited has a different problem from a page that isn't being retrieved at all. These two problems require different fixes. Most AEO audit approaches measure only final citation presence, which doesn't distinguish between the two. When citation rates stay low despite established directory presence, the retrieval-to-citation gap is the most likely cause -- and title alignment and URL readability are the first things to check.

Signal Check at sourcepull.ca shows per-platform citation rates across ChatGPT, Perplexity, Gemini, and Claude. For businesses with established directory presence but persistently low citation rates, a full audit report identifies which specific infrastructure gaps -- including page-level metadata alignment issues -- are likely responsible.