If the search interface becomes ChatGPT, Perplexity, or Google's AI Overviews instead of the ten blue links, what survives of a small editorial site's voice?
This repository is the public notebook for a research project I (Avigad Asky) am running on thedrop.gifts. The site is a working editorial publication — 345 kit-style gift guides, each anchored on a single product, each written for a specific recipient. The experiment is to publish honestly, instrument it carefully, and watch what an AI-mediated search era does to a small site.
Most "AI search" coverage focuses on the publisher-side panic — declining clickthroughs from Google as AI Overviews answer the query inside the SERP. But the actual editorial question is more interesting:
- When an LLM answers a question about gifts, what does it cite? Mostly Wirecutter, NYT, The Strategist, the big SEO winners. The mid-list independents are paraphrased, not cited.
- When a small site publishes a sharp opinion ("the YoYoFactory Shutter Wide is the right second yo-yo, the One Drop is the third"), does the LLM carry the opinion forward, or does it sand the voice down to a list of bullet points?
- Does serving a
/llms.txtfile (the llmstxt.org convention) actually change citation behaviour, or is it just SEO theatre for a different generation? - Can a small site, with discipline, rank in AI answers in a way that survives the next algorithm change?
thedrop.gifts is the corpus. Every page is one anchor product, six to eight supporting picks, and a specific recipient (the new espresso owner, the woodworker upgrading their bench, the kid who just learned to crochet). The editorial voice is deliberate — sharp, opinionated, gear-specific — because the question I want to answer is whether voice survives translation through an LLM at all.
The site is operated end-to-end by AI agents writing under a strict editorial-voice classifier. Humans review only when the classifier flags something. This is not the future of publishing — it is one specific experiment in what publishing has to look like in 2026 if you want to be cited rather than paraphrased.
A public-facing research notebook. Methodology notes, citation tracking, what's working, what isn't. The site itself is the deliverable; this repo is where I write down what I'm learning, in the open, so other small publishers can pick up whatever's useful.
Things I plan to publish here over the next several months:
- Citation-tracking methodology. When ChatGPT answers "what's a good gift for someone starting BJJ?" — what does it cite? How do you measure that without a research-grade query bank? I'll publish my query bank and the harness.
- The
/llms.txtexperiment. I shipped one. Does it change anything? I'll measure citation deltas across Perplexity, Claude, ChatGPT, and Gemini at fixed intervals and publish the results. - Voice persistence. When an LLM summarises one of my kits, how much of the editorial position survives? I'll publish a corpus of "summarise this page" outputs across models, scored against the original voice.
- What gets indexed by AI Overviews. Of the 345 kits, which appear in Google's AI Overviews for their target queries? Which don't? Why?
If you're a researcher, journalist, or curious editor: the kits are at thedrop.gifts. The full corpus is also published as a single Markdown file at /llms-full.txt if you want to feed it to a model.
A curated index lives at github.com/avigadasky/awesome-starter-kits — 315 kits organised by hobby, link-only.
Because the AI-search transition will be quiet, will mostly look like noise, and will be hard to study after the fact. Small sites either survive it or they don't, and I'd rather publish the instrument than guess.
If you're running a small editorial site and thinking about the same problem, I'd love to hear from you. avigad@thedrop.gifts.
Avigad Asky — operator, thedrop.gifts