Real AI prompts look like conversations, not SEO keywords. On episode 30 of The GEO Show, Paris Childress covers What If Web's finding that median prompts run 9–14 words with under 10% keyword-like strings, Reddit's ChatGPT citation crash after a robots.txt change, Ospia zero-citation mentions, Google's partner web search API, Meltwater's LinkedIn citation surge, Context7 agent tracking, Trade Desk verification behavior, a failed A-vs-B bias replication, and Profound Sheets Templates.
Real AI prompts look like conversations, not SEO keywords. On episode 30 of The GEO Show, Paris Childress centers What If Web research showing median self-reported prompts of 9–14 words, with fewer than 10% resembling short keyword strings, and argues that keyword volume is the wrong map for prompt demand.
Around that throughline he covers AirOps Quill, BrightEdge on Reddit's ChatGPT citation crash after a robots.txt change, Ospia zero-citation mentions, Google's partner-only full web search API, Meltwater's LinkedIn citation surge, Context7 agent-access tracking, Trade Desk/PA Consulting verification behavior, Ospia's failed A-vs-B brand-order replication, and Profound Sheets Templates.
Because people type conversations. What If Web surveyed 300 New Zealanders and collected more than 1,200 prompts in their own words. Median length was 9–14 words. Fewer than 10% looked like short keyword strings. In one survey, only 2% explicitly named a brand. Paris called this the most important story of the day and argued there will not be a clean keyword-research equivalent for prompts: the long tail is nearly infinite, so extrapolating keyword volume to prompt demand is the wrong model.
Real AI prompts look much more like conversations than SEO keywords.
BrightEdge tracked a fixed US education prompt panel and saw Reddit's ChatGPT citation presence fall from 17% to 3% to 0% over two weeks after a robots.txt change. Timing supports the access-policy mechanism without proving causation. The ChatGPT decline was steeper than in Google AI Overviews. A brand can lose visibility because a cited source or an engine changes crawl rules, not because its own content changed.
Ospia found ChatGPT returned zero citations on 12 of 20 answers yet named at least one vendor in 13 of 20. Gemini returned a median 47 source annotations; Perplexity returned 20. Models do not surface citations equally. Paris treats mentions as the GEO prize and weights them more heavily than citations in share-of-voice calculations.
Meltwater's ~7.3 million citations across eight platforms showed LinkedIn citations up 25% month over month to 100.2k, with Perplexity alone producing 69,302 LinkedIn citations, while Facebook and Instagram fell sharply. Context7 Docs7 tracks agent visits separately from human search and warns that an agent request is not a citation. Google documented a partner-only full web search API (JSON REST/gRPC, surfaced September 11), positioning classic search as programmable retrieval for approved agents.
Trade Desk Intelligence and PA Consulting report that 95% of consumers verify AI-generated search results elsewhere and are 1.6× more likely to finalize a purchase on the open internet than through AI tools. Paris ties that to missing solution links inside many answers: citations support the synthesis, but buyers often leave the chat and search the named brands. Ospia's A-vs-B replication found word order changed Google AI Overview recommendations in 0 of 16 brand pairs. Profound Sheets Templates push GEO platforms from analytics into spreadsheet-scale task-agent execution.
00:20. AirOps Quill: agent captain for complex GEO tools; Asana +93%, Parallel +165% ChatGPT citations.
02:28. BrightEdge: Reddit ChatGPT citations 17%→0% after robots.txt change.
10:34. Throughline: What If Web median prompts 9–14 words; keyword→prompt extrapolation fails.
13:02. Trade Desk / PA Consulting: 95% verify AI results; 1.6× purchase likelihood on open internet.
16:13. Profound Sheets Templates: thousands of agents in parallel for GEO execution tasks.
The throughline is prompt reality itself: until discovery models start from conversational behavior rather than keyword head terms, GEO teams risk optimizing a map that does not match how people actually ask AI systems for help.
Hi, everybody. Welcome to episode 30 of The GEO Show, brought to you by GEOforge, full self-driving for AI visibility. I'm your host, Paris Childress. Let's get right into some top stories for the day, and we'll start with AirOps.
AirOps launches Quill, an AI agent built to close AI search gaps. AirOps launched Quill, which it calls an AI agent captain designed to identify AI search gaps, execute changes, measure results, and learn from previous runs. AirOps says Asana increased ChatGPT citations 93% in just two weeks and Parallel increased citations 165%.
It's an interesting move. I think that as these tools become more and more complex, and AirOps is probably one of the most complex GEO tools out there, you really need to have this AI agent captain, who they call Quill. We have one called Veck, and it just helps users figure out where they should prioritize their time in the platform, and it gives them some strategic direction.
So rather than using the platform alone, just like the way that ChatGPT and Claude create a chat interface layer between a user and a piece of software via MCP, the same thing is happening inside of the software platforms themselves, and I think it's a very good move. AI agents like Quill or like Veck in GEOforge provide a friendlier interface when the user is not sure where to go, what to do, what to prioritize, or even how to build complex workflows.
They don't really need to do that now. They can go straight to the agent and just chat with the agent and say, "Tell me where I should be focusing. Help me build out these workflows. Point me in the right direction for success."
Next story: crawler access erased Reddit's ChatGPT visibility without changing its content. BrightEdge reported in a fixed US education prompt panel that Reddit's ChatGPT citation presence went from 17% down to 3% and then down to 0% across two weeks. This break coincided with Reddit's robots.txt change, and BrightEdge correctly says that the timing supports the mechanism but does not prove causation.
And ChatGPT's decline was even steeper, much steeper than in Google AI Overviews. So a brand can lose a lot of visibility just because one source or one engine changes its access policy. It's not a direct impact. It's a very indirect impact. But one change to Reddit's robots.txt file, which is the file that dictates the rules for how AI bots or all crawlers in general can access their website, caused a massive drop in the citation presence from Reddit inside of ChatGPT.
Moving on to the next story: a zero-citation answer can still recommend the brand. Ospia AI vendor research reported that across twenty prompts, ChatGPT returned zero citations on twelve out of twenty answers. Yet at least one vendor was named in thirteen out of those twenty. Gemini returned a median forty-seven source annotations per answer. Perplexity returned twenty. The study used sixty answers on one day in one US setup, so the exact percentages should not be generalized.
Interestingly here, we have twenty prompts where ChatGPT, for the majority of these, returned zero citations, but still was making mentions or recommending a vendor in its answer. And we did not see that behavior in Gemini or Perplexity, which returned a large number of source annotations or citations. So one takeaway is that the models do not behave the same in showing citations. High citation counts are not equal. It's very, very different. And also, a mention is very different than a citation.
And I've long argued that a mention is really the gold standard. This is the real prize in GEO right now: getting a mention, getting your brand mentioned in the answer as opposed to just being cited as a source. And it's one of the reasons why we give much more weighting to mentions in our share-of-voice calculation than we do to citations.
Next story: Google documents a partner-only API for full web search results. Google's web search service API allows approved programmatic partners to retrieve Google search results in JSON through REST or gRPC. Access requires a cloud project API key and a client ID tied to a partner agreement.
Google calls the operation a full web search, and this was surfaced in Google's support documentation on September 11th. Partner eligibility and pricing remain unclear. So here what we're seeing is Google's web index is becoming programmable infrastructure for selected applications and agents.
So I think Google is also gearing up for this agent-dominated future where it will have the ability to grant permission for certain agents to full web search. We've seen it happen first with the web, where AI bots and AI agents are taking over the public web. Well, that might also happen with search, with classic search. Search engines might really become infrastructure for AI retrieval effectively. It's pretty fascinating to see this development.
Next story: LinkedIn moves into Meltwater's top three AI citation sources. Meltwater's vendor research looked at roughly 7.3 million citations on eight AI platforms. That's a huge data set. And what they found was that LinkedIn citations rose twenty-five percent month over month to 100.2 thousand. Perplexity alone produced 69,302 LinkedIn citations. Meanwhile, earned/news citations fell from thirty-three percent to twenty-nine percent of citation domain counts.
And Meltwater also found sharp divergence within social citations. Facebook fell almost thirty percent and Instagram fell forty-three percent. What I would like to know here, especially in this research, is how much of these citations and responses were B2B intent versus B2C? Because what I'm seeing in this data is that LinkedIn made a big surge upward, especially in Perplexity, which leads me to believe that there was high B2B intent, but we don't know that from this research.
One thing is clear: LinkedIn is becoming stronger and stronger as a citation source, especially for B2B. And if you are doing B2B GEO, I would strongly recommend that you focus heavily on LinkedIn articles. What we do every day with this episode is we repurpose the GEO Show daily episode into a LinkedIn article, which we push out through our GEOforge Daily newsletter.
And one of the reasons that we do that, of course, is to build an audience on LinkedIn and to access that reach. I think we're now getting close to about 850 subscribers to that newsletter. So number one is we want that audience reach, but number two, we want to put ourselves in the position to be winning LinkedIn citations through the publishing of these articles, and this data really supports that strategy.
Next up: AI visibility is expanding from answer tracking to agent access tracking. Context7's Docs7 AI Visibility tracks visits from OpenAI, Claude, Perplexity, Cursor, OpenCode, Meta, and Mistral. It also exposes queries sent through its WebMCP SearchDocs tool separately from human search. Context7 explicitly warns that an agent request does not mean the page was cited or used in a final answer.
So here, what we're seeing is that the category is going beyond trying to answer "what did ChatGPT say about us" and more into the realm of how agents access and interrogate a company's information. That's becoming the broader theme here.
Next story: real AI prompts look much more like conversations than SEO keywords. Now, I think this is the most important story of the day, in my opinion. A research company called What If Web surveyed three hundred New Zealanders, and they collected more than twelve hundred prompts in the respondents' own words. So we have real research happening from real people now. Median prompts were nine to fourteen words long.
Fewer than ten percent of those prompts resembled short keyword strings. In one survey, only two percent explicitly named a brand. The authors appropriately described the findings as directional because the prompts were self-reported and they covered only three categories. Okay, it's a small data set, but the findings are very, very important here because a lot of prompt discovery and prompt research is built on the back of keyword research, and I think that this is the wrong approach, and this research is backing me up here.
There is no equivalent of a keyword research tool for prompts, and I don't think there ever will be, because as we've seen here, nine to fourteen words is the median prompt length. The variability in the long tail is almost infinite. So I don't think it is a correct approach to take keyword demand and keyword search volume and then extrapolate that out to prompts.
They are very, very different, and this research proves it. So if you're ever trying to figure out how much prompt demand is there for a particular topic and you think, well, let's do the keyword research on this topic and let's generally take that same volume and extrapolate that to prompt volume, I think that while that does make logical sense in theory, I think it's the wrong approach.
And we don't really know: are there more prompts in this topic category than there are keyword searches, or are there less? And I think this is just a blind spot that we're all going to have to continue to live with for much longer in this prompt discovery process. I don't think we're going to see a prompt research tool offered up by OpenAI or by Google anytime soon, even within the ad platforms.
Next story: 95% of consumers still verify AI-generated search results elsewhere. The Trade Desk Intelligence and PA Consulting reported that 95% of consumers double-check AI-generated search results, often on the open internet. Their wider study covered 3,000 US and UK consumers and reports that consumers are 1.6 times more likely to finalize a purchase on the open internet than through AI tools.
Okay, this is not surprising to me at all, really, because most of the time what you get in an AI answer is an answer that does not contain links. And if you're at the end of the journey, you've asked all your questions and got all your answers, and you're ready to transact, most likely you're going to go to what it says here is the open internet, but probably you're going to go to search, and you're going to start searching for some of those brands or solution names that were recommended to you in that chat conversation.
So 95% of consumers double-checking AI-generated search results, I believe, is a direct reflection of the lack of links in those AI-generated search results. And here we're talking about the mentions and the actual responses. We're not talking about the citations, because the citations are really source links that support the synthesized answer, but they are not necessarily links that point directly to the recommended solutions that are present inside of the answers themselves. And that is a very important distinction.
Next story: a viral brand-first bias claim failed replication. Ospia ran one hundred and two Google AI Overview queries across sixteen brand pairs, reversing A versus B into B versus A and repeating those queries. Word order changed the recommendation in zero of those sixteen pairs. The query-first brand appeared first forty-nine percent of the time, which is effectively a coin toss. So takeaway: query phrasing and the Google surface mattered more. And of course, this is a single-day vendor experiment. It's not definitive evidence.
But if there was ever any kind of a myth around A versus B versus B versus A, meaning if you're doing a comparison page of two brands that might include your brand, does it matter to put your brand first in the A position or in the B position? What this is showing us is, no, it doesn't matter at all. The engines are effectively not changing the recommendation due to the order of your A versus B format.
Let's move on to the last story from Profound. Pretty big news from Profound on the product front. They have added spreadsheet-scale bulk execution. Profound launched Sheets Templates on September 11th. Profound Sheets can run thousands of agents in parallel, and the six initial templates cover article translation, content refreshing, GEO FAQs, internal linking, content briefs, and product detail page optimization.
So Profound is showing us that they are moving aggressively from analytics into execution by deploying all of these different independent AI agents, and they're doing that in an interesting template that mimics a spreadsheet, the Sheets template. And I think that this is definitely the thing that we're trying to do as well with GEOforge. We're trying to take a complex workflow and assign very specific tasks to individual agents, and these agents all have a very narrow, specific focus on each of those tasks. And they receive finished work from one agent, and then the next agent in the chain gets to work on something else.
But they can also work in parallel. They can work independently. It really depends on the task. But this is the right way to architect a complex GEO platform, which is to map out the workflow, decide exactly all of the unique tasks that need to happen, which ones have critical dependencies, which ones can run in parallel or independently of each other, when to run which ones, and then effectively you build AI agents for every specific task.
You give them detailed instructions, and then you observe their behavior and their outputs, and you improve their quality over time through evals and through improving their prompt instructions. So good job to Profound moving in this direction. It's definitely the right direction, and it is probably one of the best ways to organize a complex product like we saw in the first story with AirOps.
All right. That will wrap up episode 30. Thank you all for listening, and we'll see you on the next one.
