Original insights by Cyrus Shepard
30-second rundown
Key learnings:
- More than 130 search experts scored possible influences on AI visibility: They rated how strongly each page, brand or technical characteristic might affect visibility in Google’s AI answers, using a scale from -3 to +3. Negative scores suggest it may reduce visibility, zero indicates little perceived influence, and positive scores indicate greater perceived influence.
- Access and relevance received the highest scores: AI crawl access and snippet eligibility scored +2.20, while a direct query-answer match scored +2.15.
- Recognition and evidence beat AI-facing gimmicks: Brand presence, citable facts, original information and cross-web corroboration all scored strongly, while llms.txt received just +0.05.
This week: Pick five commercial pages. Check whether Google can crawl the content, whether the page answers the question early, and whether it contains one specific fact worth citing. Fix the weakest page, then record the current AI result as your baseline.
Coming 3 months: Use an AI Search monitoring tool to track five important commercial questions daily across Google AI surfaces. Record whether your brand is included or cited, which competitors appear and which sources support each answer. Give one person responsibility for investigating movement and strengthening crawl access, on-page evidence or trusted third-party corroboration.
Cyrus Shepard asked more than 130 search experts what makes Google’s AI answers surface, select or cite a page. The result is not a stolen blueprint from the ranking bunker. It is a map drawn by experienced scouts who admit the weather may rearrange the desert by Tuesday.
The useful message is brutally ordinary. Google’s AI systems still need access, relevance, evidence and trust. The exotic trinkets were left rattling near the bottom of the suitcase.
1. The Gate Opens Before the Contest Begins
AI crawl access and snippet eligibility received the highest average score, +2.20 on a scale from -3 to +3. Query-answer match followed at +2.15. If Google cannot retrieve the page, or the page circles the question like a nervous coyote, there is little available to cite.
Traditional search did not vanish when the AI carnival arrived. Ranking for related fan-out queries scored +1.91, and ranking for the main organic query scored +1.89. Fan-out means Google may generate related subqueries, then retrieve supporting material for them.
The practical unit of optimization is no longer only one page against one keyword. A commercial insurance firm answering “What cover does a small restaurant need?” may also need strong passages on equipment breakdown, alcohol liability, delivery drivers and seasonal staff. Those supporting answers can become doors into the final response.
For a small team, the first move is not mystical. Check indexing, robots directives, snippet eligibility and server-rendered content. Then put a plain answer near the top. No incense. Make the useful thing reachable.
2. Google Wants a Known Witness With Receipts
Brand or entity presence in the model’s existing knowledge scored +2.08. Shepard calls this parametric memory: what the model already appears to know before fresh retrieval begins. It is difficult to influence quickly, which is precisely why last-minute optimization tricks look like counterfeit casino chips.
Citable facts scored +2.07. Unique first-party information reached +1.85, cross-web corroboration +1.81, and publisher reputation +1.78. The answer engine wants a recognizable witness, specific testimony and credible sources that do not contradict it.
Being quotable is not the same as being loud. Publish measured results, named methods, clear prices, product limitations, original observations and concrete customer outcomes. Then make sure partner pages, respected directories, journalism, reviews and professional profiles support the basic entity facts.
The tension matters. Unique claims attract attention, but unsupported claims from an unknown publisher can be ignored. A manufacturer saying its coating lasts 30% longer needs the test method, conditions and results on the page. Independent validation makes it harder to dismiss. The machine is not impressed by an adjective wearing a lab coat.
3. Structure Helps, but the Magic File Is Selling Snake Oil
Extractable structure scored +1.69 and answer prominence +1.65. Clear headings, tables, lists and complete sentences help systems lift the right passage. Structured data scored a more modest +0.80, with product and merchant data noted as a useful exception for shopping results.
Then came llms.txt at +0.05, essentially lying face-down in the parking lot. One contributor joked that the file plus eight dollars buys a decent boba. The survey’s point is not that technical hygiene is useless. It is that a special AI-facing file cannot manufacture reputation, relevance or evidence.
There is a warning stamped across the exercise. Respondents rated what they believe influences Google AI visibility. They did not manipulate each factor in controlled experiments and prove causation. Several experts said the findings may age badly.
That caveat is not a reason to freeze. It is a reason to avoid expensive superstition. Keep pages crawlable. Answer real questions directly. Publish facts competitors cannot copy. Build recognition and corroboration across the web. Measure the same queries every day, because the source mix can move without asking permission.
Google’s AI bouncer did not request a secret handshake. It checked whether the door opened, whether the answer fit, whether the town knew your name, and whether anybody trustworthy could confirm the story. Bring those four things, or enjoy your boba in the parking lot.
