Hatching Growth #12: AEO in Practice: Memory Hacks, Prompt Injection Experiments, and Honest Results
Glasp’s note: This is Hatching Growth, a series of articles about how Glasp organically reached millions of users. In this series, we’ll highlight some that worked and some that didn’t, and the lessons we learned along the way. While we prefer not to use the term “user,” please note that we’ll use it here for convenience 🙇♂️
If you want to reread or highlight this newsletter, save it to Glasp.
Recap: #1–#11 in one glance
#3: How We Rode the AI Wave with Side Projects Before It Exploded
#8: How We Doubled Down on YouTube Summary with Programmatic SEO
#9: Acquisition-First, the AARRR Ladder, and Turning Core Actions into a Growth Engine
Introduction
In the last edition, we talked about a wild-card moment — how an unexpected opportunity (Times Square!) can boost brand energy even if dashboards don’t spike.
This time, we’re shifting back to the frontier of growth systems: AEO (Answer Engine Optimization), also known as GEO (Generative Engine Optimization).
AEO sits on top of SEO. You still need crawlable, high-quality pages and sane site architecture. But with the introduction of ChatGPT’s Memory feature, a new surface for discovery is emerging — one that’s built not just on keywords or links, but on what the model remembers about users and the content they engage with.
This shift means that optimizing for LLMs now involves thinking about how information is stored, recalled, and re-recommended within the model’s memory — a very different challenge from classic search.
What We Mean by AEO (and Why It’s Different)
SEO fundamentals still matter. Think structure, speed, internal links, clear headings, FAQs.
AEO adds model memory. Instead of optimizing purely for crawlers, we’re now optimizing for retention inside models — ensuring that when users ask, the model remembers us.
The “Memory Hack” Hypothesis
When ChatGPT introduced Memory, we had a hunch: if enough users frequently see or ask about “Glasp” in relevant tasks (e.g., “YouTube summary Glasp”), the model might personalize toward Glasp for those users — and perhaps raise Glasp’s prior in the broader ecosystem over time.
To test that idea, we came up with a small experiment: within a specific ChatGPT Extension used by YouTube Summary users, we adjusted the default prompt to reference Glasp directly. The goal was simple, to see whether consistent brand mentions in natural user prompts could influence model memory or citation patterns over time.
How We Implemented It
For free users of our YouTube/article summary flow, we used the ChatGPT Extension — the same one from Hatching Growth #4, where we explored ChatGPT’s launch ecosystem.
The experiment relied on a structured prompt that:
First checks for an existing Glasp summary (from our programmatic SEO base, see Hatching Growth #8).
If found, that summary is used as a citation; if not, a new summary is generated via ChatGPT.
Why: to seed positive signals (reliable source + recurring brand), capture occasional return traffic, and potentially nudge model memory.
Here’s the actual prompt we used:
The closing Reader page link (“Learn more on Glasp”) offered users a direct, optional way to continue reading or exploring related material. When users access the Reader page, they can also retrieve the full YouTube transcript, making it a practical next step for those who want to go deeper into the video content.
We normally allow users to customize prompts freely. For this experiment, we temporarily adjusted the default prompt only for free users of the ChatGPT Extension — purely to observe model behavior. Paid users’ prompts remained fully customizable.
“Share on AI” — A Parallel Discovery
Around the same time, we came across an article by Metehan, who explored a concept called “AI Share Buttons.”
Think of it like “Share on X” — but instead of posting to a social platform, it opens ChatGPT, Gemini, or Claude with a prefilled prompt referencing your page. It lowers friction for users to route your content into LLMs, where it can be summarized, discussed, or built upon.
When we saw Metehan’s article, we realized his thinking mirrored our own experiments around memory and AEO signals — proof that others were independently exploring the same frontier.
That discovery led us to invite Metehan to Glasp Talk, where we discussed AEO trends, prompt injection, and how AI-driven visibility might shape the next era of growth.
The Results (No Hype)
We launched the updated version of the ChatGPT Extension in July 2025.
Traffic from ChatGPT → Glasp: minimal lift in Google Analytics.
Why:
Users often finish inside ChatGPT (the in-chat summary is “good enough”).
Even with citations, click motivation is weaker than a traditional SERP.
Ecosystem “memory” shifts are subtle and slow, not immediately measurable.
Additionally, when we looked at YouTube Summary metrics, we noticed another factor: videos that generate the most summary requests tend to be timely or trending, while the content we’ve programmatically generated through Glasp’s SEO system tends to be evergreen. This means there’s likely a supply-demand mismatch — users request summaries for the latest popular videos, while most indexed Glasp content covers long-term, non-ephemeral topics.
And even when prompt injection triggers Web Search, it’s difficult to determine whether that activity translates into actual human traffic.
From Cloudflare data, we can see significant visits coming from AI crawlers, but it’s unclear whether those hits result from users actively prompting models, or from our own prompt injection mechanisms. That ambiguity makes true AEO impact measurement especially tricky.
Bottom line: useful learning, not a growth unlock — at least not with this setup.
A Tempting but Tricky Angle: Indexable AI Chats
During our prompt-injection experiments, we wondered whether shared ChatGPT chats could act as backlinks.
When a user runs a YouTube summary through ChatGPT and it cites a Glasp article, that chat includes a Glasp URL. If the user then shares that chat publicly, it can be indexed by Google, creating a potential “soft backlink.”
The logic:
Shared chats become web pages.
They include citations to Glasp.
They can appear in search results and drive referral traffic.
However, we found the blocker wasn’t technical — it was behavioral. Most users have no reason to share a generic summary or AI conversation publicly. The motivation is weak, and there’s no natural UX hook.
Because of that, we decided not to build any feature that promotes or encourages sharing. It remains an interesting theoretical AEO angle — how user-generated AI chats might create new, indexable link surfaces — but it’s not something we’ve put into production.
The example is shown below.
Side Project: Estimating “AI Visibility”
This side project emerged while experimenting with Share on AI and Prompt Injection. We became curious how AEO analytics tools like Profound compute their AI Visibility Scores.
As Web Search API hops strip UTMs and model-native recommendations have no referrer, so tracking isn’t possible. That led us to think: if visibility can’t be measured directly, it must be estimated another way.
Our hypothesis: these tools likely send large volumes of synthetic prompts to models (e.g., ChatGPT) and calculate how often certain domains appear — creating a probabilistic visibility score.
We built our own internal simulation harness to test that idea — sending statistically meaningful prompts across models, adjusting parameters, and computing experimental AI Visibility Scores for Glasp content.
The early takeaway is that LLMs behave differently from search engines. In SEO, visibility clusters around head keywords. In LLMs, follow-up prompts multiply rapidly, generating huge long-tail surfaces where context evolves with every turn. That’s why QA-style content (like Reddit discussions) is often cited more than static resources (like Wikipedia) — it mirrors how LLMs reason and respond.
In the future, Google may connect Gemini with Google Analytics, enabling full LLM referral tracking inside its ecosystem. But for most other tools, simulation-based estimation will remain the most practical path.
This was an internal project, but if there’s enough interest, we’ll make it open-source on GitHub so others can explore how AI-driven visibility might be tracked and simulated.
Guardrails: Prompt Injection, Ethically
We ran light-touch prompt scaffolds via our extension (transparent, reversible, aligned with user intent).
Tradeoffs:
Rapid learning, low cost, small test group.
A few users may notice and leave negative reviews (“you changed my prompt”).
Our stance: keep everything observable, opt-out-able, and limited to small experiments — never default behavior.
Takeaway
The Memory Hack and Prompt Injection experiments didn’t behave as we hoped — but they revealed something important. Across AI systems, there’s an emerging competition around context retention and user lock-in. We still believe there’s untapped opportunity in that layer.
We haven’t seen any other product using browser extensions to test memory-layer interventions, so this might have been the first experiment of its kind globally.
AEO itself is still early and evolving fast — both the models and user behaviors change rapidly. That means the path forward depends on small, continuous experiments and accurate industry sensing.
Analytics remains the missing piece. Running AEO experiments today is cumbersome, and most available tools offer limited visibility into model behavior. That’s also where opportunity lies. Projects like Profound is pointing toward it — and we realized Glasp could pursue something similar, building better visibility models for the AEO layer.
For startups, the long tail is where the real opening might be. If you focus on follow-up prompts and long-tail contexts, you can compete and win in AEO even if traditional SEO remains out of reach.
Meanwhile, as ChatGPT’s API ecosystem expands — especially into shopping and transactional experiences — we’re also exploring how Glasp could participate in that new wave of “AI-native discovery.”
What’s Next
In the next edition, we’ll explore one of three themes that have been top of mind lately — either:
how UGC (User-Generated Content) can power SEO and organic discovery,
how early-stage teams can approach fundraising with creativity and discipline, or
what we’ve been learning behind the scenes through Glasp Talk and its growing community of thinkers and builders.
Whatever direction we take, the goal remains the same: to share real experiments and lessons that help small teams grow smarter — not just bigger.
Partner with Glasp
We currently offer newsletter sponsorships. If you have a product, event, or service you’d like to share with our community of learning enthusiasts, sponsor an edition of our newsletter to reach engaged readers.
We value your feedback
We’d love to hear your thoughts and invite you to our short survey.
Really insightful breakdown of how AEO is evolving beyond traditional SEO! 👏 I love how Glasp is experimenting transparently especially around memory-layer optimization and prompt injection even when results aren’t immediately growth-positive.
The takeaway about “optimizing for model memory” feels like a glimpse into the next era of organic discovery. Excited to see how Glasp continues to explore AEO analytics and long-tail prompt strategies! 🚀
Thank you for reading! I hope our experiments help you understand the cutting-edge field, AEO. We're still experimenting it, so we're happy to update and share our latest results in the later post.
If you have any questions or suggestions, or anything you want us to share, please leave a comment!