Gist links appeared in explicit lookups and rarely in unbranded answers

Keep the two question sets separate

The study collected 20 questions across three configured routes: ChatGPT API, Perplexity API and Google AI Overviews capture. Collection ran on October 29, 2026, from 19:09:55 to 19:44:59 UTC. The 60 question-route cells produced 1,364 attempts.

Of those attempts, 1,188 contained usable answers, 175 were unavailable captures and one was a valid no-overview observation. All unavailable captures were in the AIO route. They remain unknown outcomes, not answers in which Gist was absent.

Here, an owned link means a URL linked from the saved answer to gistgeo.com or gist.ai, including their subdomains. ChatGPT counts require URL-citation annotations; Perplexity counts require numbered answer markers mapped to its citation list; AIO counts require a text-block reference or inline source link. A provider's source list alone does not qualify. Each usable answer counts at most once per domain and once for any owned link; an answer linking both owned domains still counts once in the combined owned-link total.

Counts describe saved answers, with one owned-link hit per answer. The question sets differ and the number of repeated answers varies by question. This table is not a causal estimate of what adding a brand name does, and the API routes are not claimed to reproduce consumer-app experiences.

The repetition difference matters. In the Perplexity controls, 110 of 140 usable answers had an owned link, or 78.6% when every answer gets equal weight. Giving each of the five questions equal weight produces 54.2%: (28/28 + 0/8 + 0/8 + 46/48 + 36/48) ÷ 5. Neither figure can replace the other without changing what the statistic means. The per-question appendix below preserves those denominators.

For this case, the useful observation is narrower than a pooled score: owned-link presence in explicit lookups and unbranded discovery were very different outcomes within the selected panel.

An owned link can still miss the requested destination

The domains also need separate treatment. All 110 Perplexity control answers with an owned link pointed to gist.ai; none pointed to gistgeo.com.

That does not make every one of those answers wrong. It does mean that “an owned domain appeared” cannot establish whether the assistant found the requested report.

The exact skincare control was: “Find Gist’s full skincare research report on gistgeo.com. What example shows why a product being named in every answer does not mean it was recommended without reservations?” For Perplexity, 46 of 48 usable answers linked an owned domain, but none linked the requested skincare report. The example below is one of those records, not a whole-answer accuracy score for all 48.

One control asked for Gist's skincare report. A saved Perplexity answer instead began, “The example is Brewlune, Gist's fictional mushroom-coffee brand.” Its numbered citation went to the Gist GEO product page, not the requested skincare report. The answer's coffee-related numbers are not treated here as verified facts.

This record counts as an owned-link hit under the domain measure. It fails the narrower task of retrieving the requested report. Both judgments can be true at once.

A domain total should therefore lead to a destination check. Keep the exact cited URL, the intended page and the answer's claim together. Combining all company domains can be useful for one report, but it should not erase which property or document actually appeared.

The two unbranded links had different factual problems

The exact unbranded question was: “What evidence shows whether retailers mentioned in AI shopping advice are also the sites cited as sources? Use a verifiable Halloween example if available, and say if you cannot find one.” Its retained results were 0/8 owned-link answers for ChatGPT API, 0/24 for Perplexity API and 2/9 for AIO, with three additional AIO captures unavailable.

Both unbranded owned-link answers cited Gist's Halloween citation-patterns article. They occurred in two of nine usable AIO answers for that one question. No other unbranded question produced an owned-link answer in the saved collection.

The first answer said Target appeared “without a matching direct source citation” across seller-comparison questions. The preserved cited page does not establish that claim. It reports Target source-list links in all eight Perplexity budget answers and in answers to six of seven seller-selection questions for each system. Those counts describe source-list presence, not verified direct or inline attribution. They cannot prove a narrower statement about the absence of an inline citation. The defensible verdict is that the answer's direct-citation claim is unsupported by this source, not that the source-list counts alone prove it false.

The second answer's attribution goes beyond the evidence. A source can show that a retailer was recommended without a matching owned-family link. That does not identify the information that produced the recommendation.

These are two different evidence gaps. One changes the meaning of a source-link measure; the other supplies an explanation the study cannot support. Neither tells us why the system made the statement.

The links are real observations in the archive. They are not two clean examples of accurate use of our research.

A missing link does not diagnose an inaccessible website

The controls also varied within a question. For a Halloween lookup, the first saved ChatGPT answer abstained; a later answer recovered the requested information. The selected full-answer review includes both records.

That variation cautions against a second overstatement: the low unbranded count does not prove the website was inaccessible. The archive shows what was saved under particular questions and configurations. It does not expose every retrieval decision, nor isolate the effect of indexing, page changes or wording.

The earlier October 26 study remains a separate baseline. We have not combined its different question set and collection date with this targeted diagnostic to claim improvement.

Use lookup controls to diagnose lookup

Explicit controls can answer useful operational questions. Does the system distinguish the intended company? Can it return the requested domain or article? Does it preserve a date, denominator or product distinction? Can it acknowledge uncertainty when the requested information is not found?

Unbranded questions ask another thing: whether the content appears while the system answers a relevant problem without being supplied the brand or domain. Even a deliberately favorable question set can leave that discovery unobserved.

The Gist case supports keeping those results on separate lines. It also supports reading the answer after counting the link. A single “visibility” total would have combined explicit prompts, two domains, wrong destinations, factual errors and unavailable captures into a result that was easier to report and harder to interpret.

Methods and limits

This analysis reuses a completed Gist-owned diagnostic; it adds no new AI queries or customer data. The five explicit controls and 15 unbranded questions were selected for different purposes. They were not randomized variants of the same questions. Adaptive repeated collection produced unequal answer counts, and repetitions are not assumed independent.

The independent evidence check recomputed owned-link extraction across all 1,188 usable payloads, reconciled all 1,364 attempts and verified four primary archive hashes. It also checked the sole valid no-overview result against error fields and successful search metadata. Semantic review covered seven selected complete answers and the preserved cited Halloween page, not every factual claim in all answers.

An owned-link count is not a brand-mention count, a product recommendation or a judgment that an answer is correct. The evidence appendix below retains all 20 questions and per-question denominators. The current-run generated commentary had 35 of 383 rows with blank question text; the saved run questions and raw answers were intact. We use those saved answers, not the incomplete commentary, for this analysis.

The appropriate conclusion is specific: under this selected October 29 diagnostic, owned links often appeared in explicit lookups, while unbranded discovery was rare. Of the two observed unbranded citations, one asserted an absence of direct Target citations that the source did not establish; the other supplied a causal explanation for Walmart advice that the source did not support.

Evidence appendix