<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Spyglasses Research</title><description>Studies of how AI assistants find, evaluate, and cite sources — from the team behind the Spyglasses AI visibility platform.</description><link>https://research.spyglasses.io/</link><item><title>The AI shortlist held for a month</title><link>https://research.spyglasses.io/articles/ai-shortlist-stability-over-time/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/ai-shortlist-stability-over-time/</guid><description>Experiment 010, Stage 2, pre-registered: the same 40 B2B software buying questions asked on 11 days over five weeks on ChatGPT, Gemini and the Claude API. A vendor shortlisted in one answer was shortlisted again a month later about as often as a day or two later: 81% against 82% on ChatGPT and 76% against 79% on Claude. Gemini was lower and its test was inconclusive, but it showed no drift. Only Gemini&apos;s top pick clearly moved.</description><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate></item><item><title>Claude Code on a subscription names the same brands as claude.ai</title><link>https://research.spyglasses.io/articles/claude-code-vs-claude-ai/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/claude-code-vs-claude-ai/</guid><description>Experiment 011, pre-registered: 40 B2B software buying prompts on three days, typed by hand into claude.ai and run unattended through Claude Code on a Claude subscription. A scheduled Claude Code task named the same brands as claude.ai, within claude.ai&apos;s own day-to-day variation. A local Claude Code run came close, inside our equivalence band. Not suitable when search location matters.</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate></item><item><title>AI changes its sources far more than its shortlist</title><link>https://research.spyglasses.io/articles/ai-sources-vs-shortlist/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/ai-sources-vs-shortlist/</guid><description>Research note, exploratory: in 918 B2B software answers, a vendor on the AI shortlist repeated in another run 64% to 81% of the time, while third-party cited sites repeated 19% to 46% of the time. When the sources changed, the shortlist barely moved.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate></item><item><title>When AI shortlists your software, it usually shortlists you again</title><link>https://research.spyglasses.io/articles/ai-shortlist-stability/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/ai-shortlist-stability/</guid><description>Experiment 010, Stage 1, pre-registered: 40 B2B software buying questions asked on 7 days on ChatGPT, Gemini and Claude. A vendor on the AI&apos;s shortlist was shortlisted again in another run 81% of the time on ChatGPT, 78% on Claude and 64% on Gemini, and claude.ai replicated at 75%. Most of each shortlist is a stable core of vendors, while the platforms disagree with each other.</description><pubDate>Sun, 11 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Sonnet is not a stand-in for claude.ai. Opus 5.5 through the API is close.</title><link>https://research.spyglasses.io/articles/claude-api-vs-claude-ai/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/claude-api-vs-claude-ai/</guid><description>Experiment 009, pre-registered: 40 B2B software buying prompts asked on three days in claude.ai, which now defaults to Opus 5.5, and through seven Claude API configurations. Every Sonnet 5 and Haiku 4.5 configuration named a measurably different set of brands from claude.ai. Opus 5.5 through the API came within our equivalence band, and adding the leaked claude.ai system prompt raised its estimated cost by about 80% without a meaningful gain in which brands it named. The claude.ai reasoning level did not change which brands were named.</description><pubDate>Fri, 02 Oct 2026 00:00:00 GMT</pubDate></item><item><title>ChatGPT knows your domain. It is not guessing.</title><link>https://research.spyglasses.io/articles/brand-domain-knowledge/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/brand-domain-knowledge/</guid><description>Experiment 008, pre-registered: 48 brands in four tiers, asked about daily for ten days through the OpenAI Responses API with web search, recording the site: operator the model types. When it consulted a brand&apos;s own site it used the canonical domain in 98.5% to 100% of calls in every tier, including the tier whose domains cannot be derived from the name. A brand that changed domains still gets its old one consulted alongside the new one.</description><pubDate>Fri, 11 Sep 2026 00:00:00 GMT</pubDate></item><item><title>When AI consults a website directly, it is usually not the brand you asked about. The sites it picks form a measurable authority club.</title><link>https://research.spyglasses.io/articles/where-ai-looks-first/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/where-ai-looks-first/</guid><description>Experiment 007: we classified the sites ChatGPT consults directly with site:-scoped grounding searches. 81% of direct consultations go to third parties. The list leans institutional even after correcting for panel mix. Consulted domains sit far above merely-cited ones on web-graph centrality and AI Placement Value, and comparison prompts, not brand-identity prompts, are the trigger.</description><pubDate>Tue, 01 Sep 2026 00:00:00 GMT</pubDate></item><item><title>We tried sourcing prompts from real human conversations. The material is abundant. Using it is the hard part.</title><link>https://research.spyglasses.io/articles/sourcing-prompts-from-forums/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/sourcing-prompts-from-forums/</guid><description>A methodology note, not a study. The advice to source AI-visibility prompts from Reddit and similar venues sounds like it dissolves the synthetic-prompt problem. We ran it properly for one product category: volume was never the obstacle, but only about 1 post in 80 is usable as a prompt unedited. Clipping them at sentence boundaries fixes that for 93% and strips the author&apos;s sub-intent from most of them. Includes the wrong answer we reached first, and why our own frozen protocol did not catch it.</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate></item><item><title>A synthetic panel matched to human sub-intent reproduces the consideration set — but not the share of voice.</title><link>https://research.spyglasses.io/articles/subintent-matched-panels/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/subintent-matched-panels/</guid><description>Experiment 005, pre-registered: a scenario-generated prompt panel stratified to 143 human survey phrasings&apos; sub-intent profile, run beside an unstratified panel and the re-run human panel on ChatGPT for five days. Its answers are statistically indistinguishable from human ones — the closest match we have measured — and its brand-share vector still misses by 8.9 points against a 5-point band. The gap turns out to be about how often each brand appears, not which brands appear. We also failed to replicate our own finding that synthetic prompts are noisier.</description><pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate></item><item><title>We tested our own prompt generator against 143 humans. It measures your home field, and we found what a market-matching panel needs.</title><link>https://research.spyglasses.io/articles/synthetic-prompt-coverage/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/synthetic-prompt-coverage/</guid><description>Experiment 003: Four prompt panels, 143 human survey phrasings, two brand-anchored panels from our own production generator, and one neutral scenario panel — run side by side on ChatGPT for five days, pre-registered. Brand-anchored panels fail every market-representativeness test; what they measure instead is your standing on the buying situations your own positioning claims. But a neutral scenario panel&apos;s prompts are statistically indistinguishable from human phrasings, and when synthetic and human prompts share a sub-intent, their answers match. This is a concrete, testable path to synthetic panels that mirror human ones, which we are pre-registering next.</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate></item><item><title>A reader asked if &apos;rank is unstable&apos; was measured or folklore. So we measured it.</title><link>https://research.spyglasses.io/articles/rank-stability-chatgpt/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/rank-stability-chatgpt/</guid><description>Follow-up to our SparkToro replication: the first public coefficient on AI recommendation rank stability. Across 1,001 ChatGPT runs, the order of recommended brands agrees at Kendall&apos;s tau 0.64 when the same prompt is re-run and 0.36 across different phrasings — and nearly half of same-prompt pairs return shared brands in exactly the same order. The &apos;rank chaos&apos; folklore mostly isn&apos;t about rank.</description><pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate></item><item><title>We pre-registered &apos;prompt phrasing barely matters.&apos; ChatGPT proved us wrong.</title><link>https://research.spyglasses.io/articles/prompt-phrasing-consistency/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/prompt-phrasing-consistency/</guid><description>Extending SparkToro&apos;s prompt-diversity study: we pre-registered the hypothesis that prompt wording doesn&apos;t matter when intent matches, then re-ran their 143 de-identified survey prompts against one platform — ChatGPT — every day for a week. The leading brands held, but rewording moved the shortlist, the cited sources, and the searches underneath, well beyond run-to-run noise.</description><pubDate>Thu, 23 Jul 2026 00:00:00 GMT</pubDate></item><item><title>Only one AI surface deep-links to the moment inside a YouTube video</title><link>https://research.spyglasses.io/articles/youtube-moment-citations/</link><guid isPermaLink="true">https://research.spyglasses.io/articles/youtube-moment-citations/</guid><description>Google AI Overviews attaches a timestamp to 44% of the YouTube videos it cites, sending users to the exact second that answers their question. ChatGPT, Gemini, Perplexity, and Claude never do. We looked at what predicts a moment citation — and whether Google is reading the transcript.</description><pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate></item></channel></rss>