Model: ChatGPT · OpenAI

ChatGPT can now pull context from your past chats, saved memories, files and Gmail - and it shows users exactly which sources shaped every answer. For brands tracking AI visibility, the rules just changed.

Source: ChatGPT release notes - 5 May 2026

On 5 May 2026, OpenAI rolled out one of the most consequential ChatGPT updates of the year for anyone monitoring how AI platforms talk about their brand. The change does two things at once: it widens the pool of personal data ChatGPT draws on when generating a response, and it gives users a transparent view of which sources were used.

If your team builds content with the assumption that ChatGPT replies from a single, shared pool of training data, that mental model is officially out of date. Recommendations are now stitched together from a user's own history - and the brands that win in that environment will look very different from the ones that win in classic SEO.

Here is what changed, why it matters for AI visibility, and what to start tracking.

What OpenAI actually shipped on 5 May 2026

The official release notes describe two linked features that landed together:

  • Expanded memory recall. ChatGPT can now pull relevant context from past chats, saved memories and, where available, files and a connected Gmail account. OpenAI describes the search across past conversations as faster, so users do not have to repeat themselves as often.
  • Memory sources surfaced in the UI. Across every consumer plan, a Sources icon now appears alongside responses. Tapping it reveals which saved memories, past chats and custom instructions contributed to the answer. Plus and Pro subscribers additionally see references pulled from their file library and Gmail.

The update launched first on the web, with mobile following shortly after. Importantly, OpenAI clarified that memory sources are visible only inside the user's own account experience: if a chat is shared publicly, the sources behind the answer are not. Users can also delete sources, disconnect apps, switch to temporary chats, or turn memory off entirely.

Why this matters for AI visibility tracking

For the past two years, AI visibility work has loosely mirrored SEO: figure out how a model represents your brand in a given prompt, track how that representation shifts over time, and try to influence the inputs the model leans on. That framework still works - but only for the part of the answer that is not personalised.

With memory sources now active on every consumer plan, a non-trivial share of every ChatGPT response is being assembled from the user's own history. Two users in the same country, asking the same question, on the same model, can now receive meaningfully different brand recommendations - because their saved memories, prior chats and (on Plus and Pro) emails diverge.

That has three direct implications for visibility work:

1. Aggregate visibility scores are no longer a complete picture

If your dashboard shows "ChatGPT mentioned us in 38% of relevant prompts this week," that figure is still valuable - but it is now a population-level average around an increasingly personalised distribution. Heavy ChatGPT users with months of accumulated memory will see brand picks that drift away from the baseline. Your real-world coverage may be wider or narrower than the average suggests, depending on which user segments your brand has historically lived inside.

2. Repeat exposure compounds (and so does its opposite)

Once a user has interacted with your brand inside ChatGPT - asking about pricing, comparing you to a competitor, planning a purchase - that conversation becomes a memory the model can reach back to. Future answers in the same category are more likely to surface you, because you are now in the user's "relevant context" pool.

The reverse is equally true. If a user has spent six months asking about a competitor and has never mentioned you, the model will keep building on that context. The first-mover advantage inside a user's chat history is now a measurable moat.

3. Gmail and file context expand the surface area dramatically

For Plus and Pro users, ChatGPT can now reach into Gmail and uploaded files to inform its answers. That means a transactional email confirming a purchase, a contract PDF, or a newsletter sitting in someone's inbox can all become "sources" behind a recommendation.

Brands that already invest in clean transactional emails, well-structured PDFs and consistent product naming have an underrated asset here. The signal lives in the user's account, not on the open web - which means it is not something most AI visibility tools can see from the outside.

The transparency layer: a quiet win for brand auditing

The second half of the update - the Sources icon - is the part that most teams will underestimate.

For the first time, individual ChatGPT users can see, in plain English, which past conversations, files or emails shaped the answer they just received. That is a significant accountability shift. Users can now:

  • Spot when an outdated preference is still steering recommendations.
  • Delete or mark a memory as no longer relevant.
  • Disconnect Gmail or a file library if they want a more neutral answer.
  • Switch to a temporary chat that does not write back to memory.

For brands, that means a user who notices "this answer is shaped by my purchase from a competitor 18 months ago" has a one-click route to clear it. The half-life of a stale negative memory just got shorter. So did the half-life of a positive one - which is the more uncomfortable side of the equation.

What changes for Free vs Plus and Pro users

The tier split is worth understanding because it changes how different audiences will see your brand inside ChatGPT.

  • Free users: Memory sources show past chats, saved memories and custom instructions. Answers are personalised from a user's ChatGPT history but not from their wider digital life.
  • Plus and Pro users: Everything Free users get, plus references from the file library and connected Gmail account. Answers can incorporate documents, receipts, contracts and email threads.

If your buyers, evaluators or analysts skew toward Plus and Pro - and in B2B categories they almost certainly do - the Gmail and file integration is where most of the personalisation pressure will land. Expect category leaders inside someone's inbox to start dominating recommendations for that user, regardless of overall search share.

What to start tracking now

The right response to this update is not to chase the personalisation layer directly - you cannot see inside another user's chats or inbox. The response is to widen what you measure on top of your existing AI visibility tracking.

  • Baseline (cold) visibility: Keep tracking how ChatGPT represents your brand from a clean session with no memory. This is your floor - the answer a brand-new user would get.
  • Variance across modes and regions: The widening personalisation layer makes regional and mode-level differences (Free vs Plus, UK vs US) more informative. Watch them weekly, not monthly.
  • Email and transactional touchpoints: If Gmail is now a source, the consistency and structure of your transactional emails matters in a new way. Audit how your brand name, product names and pricing appear across confirmations, receipts and newsletters.
  • Memory-friendly content: Content that a user is likely to save, paste, or return to is now a multiplier. A clear comparison table or a single-page brief is more likely to become a "memory" the model leans on later than a long blog post.
  • Citation health on the open web: When a user has no relevant memory, ChatGPT falls back to its trained and retrieved sources. That fallback is still the foundation, and it is still where most visibility tools can give you a defensible signal.

The bigger picture: AI visibility is becoming two problems, not one

The May 2026 memory sources update makes a split that has been forming for a while explicit. There is now a cold visibility problem - how an AI model represents your brand to a stranger - and a personalised visibility problem - how that same model represents you to someone who already has a relationship with your category.

Most brands have been quietly solving the first problem and assuming the second will follow. After 5 May 2026, that assumption breaks. The first problem is fought in retrievable sources, citations and structured content on the open web. The second is fought in product experiences, transactional emails, and the everyday touchpoints that quietly write themselves into a user's ChatGPT memory.

The brands that build for both will compound. The brands that only optimise for the cold layer will see their aggregate metrics hold steady while their share of real, repeat-user conversations slowly erodes.

Track how ChatGPT is talking about your brand

If you want to see how ChatGPT, Gemini, Perplexity and Google AI Overview are representing your brand week over week - and how that picture is moving as OpenAI ships changes like this one - reconnAI's AI Visibility Tracking platform was built for exactly this kind of moment.

To map out how the memory sources update is likely to affect your category specifically, get in touch with our team. We will walk you through what to baseline now, before the personalised layer hardens.