How to Optimize for ChatGPT Projects and Custom GPTs

ChatGPT Projects and Custom GPTs are persistent, context-loaded workspaces where 20% of enterprise messages are already processed. Brands that appear in the knowledge files, get cited by embedded instructions, or rank in the GPT Store's discovery surfaces hold outsized influence. This guide shows exactly how to earn that position.

ChatGPT crossed 900 million weekly active users in February 2026, according to OpenAI. That number is notable on its own, but the more consequential shift is structural: a growing portion of those sessions are not one-off queries. They are persistent, file-loaded, instruction-driven workspaces, and the brands those workspaces reference are not chosen at random.

Custom GPTs usage grew 19x in 2025. Today, roughly 20% of all enterprise ChatGPT messages flow through either a Custom GPT or a Project, per OpenAI's own usage data. BBVA alone runs more than 4,000 internal Custom GPTs. The GPT Store hosts over 3 million GPTs, with approximately 159,000 publicly listed. When a user loads a Custom GPT to do competitor research, evaluate vendors, or draft a recommendation memo, what is in that GPT's knowledge base is the brand shortlist, not the open web.

If you are still treating GEO as pure content optimization for open-web queries, you are optimizing for the smaller surface. This guide covers the full picture.

What ChatGPT Projects and Custom GPTs Actually Are (and Why They Differ)

Understanding the technical distinction between these two features is prerequisite to optimizing for either.

ChatGPT Projects are persistent workspaces within the native ChatGPT interface. Each Project contains a system instruction layer, uploaded reference files (up to 40 depending on subscription tier), and a Project Memory scope that captures context across sessions within that workspace. Critically, memory is partitioned: facts learned inside a Project do not bleed into main chat or other Projects. This makes Projects ideal for multi-phase workstreams where context needs to stay anchored.

Custom GPTs are standalone AI agents built through OpenAI's GPT Builder. They combine a system prompt, uploaded knowledge files, optional web browsing and code interpreter tools, and API actions. Builders can publish them publicly in the GPT Store, share via direct link, or restrict access to their organization. Unlike Projects, Custom GPTs can be discovered by strangers through the Store, which creates an entirely different GEO surface.

The practical implication: brand mentions inside a Custom GPT's knowledge files get surfaced repeatedly, to every user who opens that GPT, often without any retrieval from the live web. It is a form of pre-retrieved citation, baked in at build time.

Why This Surface Matters for Brand Visibility

The May 7, 2026 ChatGPT update, widely called the Branded Link Update, pushed clickable brand URLs inline into AI responses at scale. The share of ChatGPT responses containing a URL jumped from roughly 4.5% to 20-24% overnight, a near-5x increase. Homepage referrals to tracked brand sites surged 354.7% week-over-week in the days that followed.

ChatGPT referral traffic now converts at 7.1%, just behind paid search at 7.8% and ahead of organic, social, and email, according to Similarweb clickstream data from April-May 2026. This is not search traffic that happens to come from AI. It is high-intent, research-stage traffic from buyers who were already in a context-loaded session when they encountered your brand.

Custom GPTs and Projects accelerate that dynamic further. A procurement analyst using a Custom GPT configured to evaluate B2B SaaS vendors is not browsing the open web. They are querying a system that was pre-loaded with a curated shortlist. The question for your brand is whether you are in that shortlist at build time, and whether open-web signals make you the obvious answer when the GPT does go to retrieve live data.

For a full view of how these citation dynamics play out across AI models, see the 2026 state of AI citations.

The Four Optimization Surfaces in the ChatGPT Ecosystem

*The ChatGPT ecosystem contains four distinct brand visibility surfaces, each requiring a different optimization lever.*

Most GEO strategies address only Surface 1. The following sections cover the remaining three.

How to Optimize for Custom GPT Knowledge Files

When a builder uploads files to a Custom GPT's knowledge base, the model retrieves from those files first, before it goes to the web (if web browsing is even enabled). Getting your brand into those files is the most durable citation position available in the ChatGPT ecosystem.

Create assets designed to be uploaded

Your goal is to produce documents that are so useful, specific, and well-structured that researchers, consultants, and practitioners choose to upload them into their own Custom GPTs. This is not metaphorical. Industry comparison guides, vendor evaluation frameworks, benchmark reports, and technical spec sheets are exactly what people load into research GPTs.

Concrete formats that get uploaded:

• Comparison tables with specific, numeric criteria across 5-8 vendors (including your own, with honest positioning)
• One-page technical summaries or product spec PDFs with clean structured headers
• Case study documents formatted as narrative + data table, not slideware
• Industry benchmark reports with original statistics, clearly attributed

The Princeton GEO study (KDD 2024) found that adding statistics increases AI citation rates by 41%, and citing sources increases them by up to 115% for lower-ranked pages. Those findings apply doubly to documents that are directly uploaded, because the model has no competing signal to override your content.

Structure for retrieval, not human reading

Custom GPTs process knowledge files using chunked retrieval. A document without clear headers is a retrieval liability: the model may surface the wrong chunk or miss your key claims entirely. Every knowledge document you create should use:

• Explicit H1/H2 header hierarchy
• Short, declarative sentences (one claim per sentence)
• Numeric specifics over descriptive adjectives
• A brief summary section at the top that front-loads your most important claims

File format matters. Plain text and Markdown files chunk more cleanly than PDFs. If you must use PDF (for branded output), ensure it is not image-rendered and that the underlying text is machine-readable.

Seed the ecosystem via third-party channels

You cannot control what files a Custom GPT builder chooses to upload, but you can influence the pool of available documents. Press coverage, analyst reports, G2/Capterra profile exports, and detailed partner case studies are all document types that practitioners pull into research GPTs. Building a content hub that AI engines cite is prerequisite to any Custom GPT strategy, because those are the assets that circulate.

How to Optimize for ChatGPT Projects

Projects differ from Custom GPTs in a key way: the person uploading files is the end user, not a third-party builder. This means your optimization lever is controlling what users have access to when they build their own research or purchasing Project.

Make your content obvious to download and reference

When a buyer starts a ChatGPT Project to evaluate solutions in your category, they are going to upload whatever they find in the first ten minutes of research. This includes your website's key pages, your downloadable resources, and anything that appears authoritative when they search your category.

Practical steps:

• Publish a dedicated "Resources for Evaluators" page or a consolidated product overview PDF, structured explicitly for people doing vendor evaluation
• Ensure your pricing page, integration documentation, and comparison content are indexable, fast-loading, and summarizable (short paragraphs, headers, no JS-gated content)
• Write case studies that include a structured summary box at the top: use case, industry, measurable outcome, timeline

Optimize for Project Memory capture

When a user runs sessions in a Project and ChatGPT's memory captures key facts from those sessions, the model is extracting claims from the content it has processed. If your content appears in that Project (either via upload or web retrieval), the facts it contains can become part of the persistent context across that user's future sessions.

This is not a technically exploitable mechanism you can game, but it is sensitive to clarity. Vague, marketing-heavy copy does not get extracted into memory. Specific, factual, structured claims do. For example: "Promptive tracks AI citation share across ChatGPT, Perplexity, Claude, and Gemini with daily prompt refresh" is the kind of claim a memory layer captures. "We help brands win in AI search" is not.

For the technical foundations that make your site parseable inside these environments, see Technical GEO: Making Your Site Crawlable and Parseable for AI Engines.

GPT Store Discoverability: A Separate Brand Channel

The GPT Store is not a citation surface in the traditional GEO sense. It is closer to an app store, with leaderboards, categories, and search. Over 3 million Custom GPTs have been created; roughly 159,000 are public and active.

Build a GPT around your category, not your brand

The highest-discoverability GPTs in the Store are utility-first: a competitor research GPT, a SEO brief generator, a vendor comparison assistant. GPTs named "[Brand] Assistant" rank primarily on brand searches. GPTs named "B2B SaaS Vendor Evaluator" or "Marketing Analytics Audit Bot" appear in category and keyword searches conducted by buyers who do not yet know your brand exists.

Build a GPT that does something genuinely useful in your category. Embed your brand lightly in the instructions, the knowledge base, and the output templates, not in the name. This is the GPT Store equivalent of writing a high-value SEO article that ranks for a non-brand query and converts to brand recognition.

Optimize your GPT listing for Store search

When a builder publishes a GPT to the Store, they set a name, description, and category. These fields behave like title tags and meta descriptions: they are the primary signal for Store search. Specific, descriptive names outperform branded or vague ones. A description that includes the category, use case, and target user type ranks better and converts browsers to users more reliably.

OpenAI also factors usage volume and ratings into the Store leaderboard. A GPT with 5,000 active users will surface above a technically identical GPT with 50. Promotion strategy (linking to your GPT from your website, including it in email onboarding sequences, embedding it in blog posts) is therefore directly tied to Store discoverability.

Schema and Structured Data for the ChatGPT Ecosystem

The Princeton GEO findings and subsequent 2026 research consistently show that structured, attributable content outperforms generic prose in AI retrieval. Content with complete Organization, Article, and Product schema is approximately 2.5 times more likely to appear in AI-generated answers than pages relying on basic Open Graph tags alone, based on multiple 2026 structured data analyses tracking AI citation rates.

The schema types that matter most in 2026:

• Organization with complete @id, founder, founding date, address, sameAs (social profiles, Wikidata), and logo, anchors your brand entity in the model's internal knowledge graph
• Article / BlogPosting with structured author, datePublished, dateModified, about, and mentions signals freshness and factual attribution
• FAQPage schema is still valid and useful for AI extraction even though Google removed FAQ rich results from SERPs in May 2026, because the schema aids model parsing, not SERP rendering
• Product with price range, features, and competitive positioning helps ChatGPT Shopping surfaces, which operate on structured product data

Only about 12% of all registered domains implement schema.org structured data, and far fewer implement it with entity relationships and @graph structure. This gap is an advantage for early movers. Run a GEO audit to identify what is missing in your current structured data implementation.

How AI Engines Cite Differently: What It Means for This Strategy

Not every AI engine uses Custom GPTs or Projects, but citation behavior across models shapes what ends up in those knowledge files. Researchers tend to pull from the sources that AI engines surface most often, so open-web citation performance and Custom GPT optimization are interconnected, not separate strategies.

Reddit is cited in roughly 40% of AI model responses across engines, and roughly 12% of ChatGPT US responses specifically. Wikipedia accounts for about 13% of ChatGPT citations. These are not destinations you can control directly, but they are surfaces where brand presence compounds. An honest, informative presence on Reddit threads and in Wikipedia categories relevant to your industry feeds into both open-web retrieval and the type of third-party documents that get uploaded into Custom GPTs.

For a full breakdown of how citation behavior differs across ChatGPT, Perplexity, Claude, and Gemini, see how AI engines cite brands differently.

The Execution Checklist

Use this checklist to track optimization across all four ChatGPT surfaces.

Content and Asset Layer

• [ ] Publish a standalone vendor evaluation guide or category comparison document in plain text or clean PDF
• [ ] Ensure all long-form content uses explicit H2/H3 headers, short paragraphs, and numeric specifics
• [ ] Create a downloadable product spec or one-pager formatted for file upload (not image-based PDF)
• [ ] Build at least one original data asset (benchmark, survey, analysis) with clear attribution

Technical Layer

• [ ] Implement Organization schema with @id, sameAs, founder, founding date, and social profiles
• [ ] Implement FAQPage schema on key category and comparison pages
• [ ] Implement Article/BlogPosting with author entity, datePublished, and dateModified on all posts
• [ ] Confirm all key pages are server-rendered (no JS gating on critical content)

GPT Store Layer

• [ ] Publish at least one utility-first Custom GPT in your category with keyword-rich name and description
• [ ] Upload brand knowledge files (comparison guide, spec sheet, case studies) to that GPT
• [ ] Link to the GPT from your website, onboarding emails, and relevant blog content
• [ ] Enable web browsing in the GPT if real-time data is relevant to its use case

Monitoring Layer

• [ ] Track ChatGPT citation share across at least 15-20 category prompts monthly
• [ ] Monitor GPT Store ranking for your published GPTs quarterly
• [ ] Audit which third-party sources in your category appear most often in ChatGPT responses, and pursue coverage there

For a 30-day structured sprint to build this foundation, follow the 30-day GEO sprint framework.

Comparison: ChatGPT Optimization Surfaces at a Glance

*Mapping each ChatGPT surface by how much control your brand has over the citation and how broadly it reaches potential buyers.*

Frequently Asked Questions

Does optimizing for Custom GPTs require me to build my own GPT?

No, though building one is a valid tactic for the GPT Store surface. The higher-leverage move is producing content, data pages, and structured documents that third-party builders choose to upload into their own GPTs. You do not need access to a builder's workspace; you need to be the most useful source in your category.

Can I see which Custom GPTs mention or reference my brand?

Not directly through OpenAI's tools. You can monitor open-web signals (review site mentions, forum discussions, third-party comparisons) that feed into GPT knowledge bases, and track ChatGPT citation share across category prompts using a GEO monitoring platform like Promptive. Direct Custom GPT attribution requires user self-report or enterprise-level OpenAI API telemetry.

Do the same content tactics work for Projects as for open-web ChatGPT retrieval?

Partially. Structured, specific, stats-heavy content performs well in both contexts. The key difference is that Projects rely on user-uploaded files, so your optimization target is making your content the obvious thing to download and reference, not just the thing ChatGPT retrieves on its own. Downloadable formats (clean PDFs, structured guides) matter more here than they do for open-web retrieval.

How does FAQ schema affect ChatGPT responses if Google removed FAQ rich results?

Google dropped FAQ rich results from SERPs in May 2026, but FAQPage schema still serves a different function: it signals question-answer structure to AI parsing systems, including ChatGPT's retrieval pipeline. The schema tells the model exactly where to find a direct answer to a specific question. There is no SERP rendering benefit, but the citation value for AI engines remains.

What file types work best in Custom GPT knowledge bases?

Plain text (.txt) and Markdown (.md) files chunk most cleanly and produce the most reliable retrieval. Clean, text-layer PDFs work well when formatted correctly. Avoid image-based PDFs, PowerPoints, or heavily designed layouts where the text is not machine-readable. Keep individual documents focused, ideally under 10 pages, and use explicit headers to guide retrieval.

How do I track whether my GEO work is improving my position in ChatGPT responses?

Run a consistent set of 15-25 category and comparison prompts through ChatGPT monthly and record whether your brand appears, in what position, and with what sentiment. A GEO monitoring platform automates this at scale, refreshes prompts daily, and surfaces share-of-voice trends across models. Manual tracking works for a pilot; it does not scale past the first 90 days.

Should I enable web browsing in my Custom GPT or rely solely on knowledge files?

Both serve different needs. Knowledge files give you deterministic, pre-loaded context that does not depend on live crawl quality. Web browsing allows the GPT to retrieve current information, which is valuable for real-time category comparisons or pricing lookups. For brand positioning purposes, enabling web browsing alongside strong knowledge files is the most defensible setup: it layers your pre-loaded content with your open-web authority signals.