A buyer in your category opened ChatGPT last Tuesday. They asked which vendor to pick. They got three names, and yours was not one of them. You never saw the session, the click, or the buyer. That is dark funnel AI search in one sentence, and it is deciding deals you will never trace.
Take a breath. This is not a mystery. It is a map you have not drawn yet.
Here is the shift worth sitting with. The dark funnel used to be a measurement headache: peer DMs, Slack threads, review-site browsing, podcast mentions, none of it showing up in your CRM. Those same surfaces are now the raw material answer engines read when a buyer asks a buying question. The untracked conversation is not just unmeasured anymore. It is the input.
By the end of this guide you will have a seven-step system: how to find the off-domain surfaces AI samples in your category, how to influence them without astroturfing, and how to measure the movement even though pixel-perfect attribution is off the table.
You do not need a bigger team for this. You need a smaller first step, run consistently.
Start by seeing the chain you are trying to influence
The chain is short, and every link happens somewhere you do not own.
A practitioner talks about a problem in a private Slack or a DM. That conversation surfaces publicly as a Reddit answer, a G2 review, or a LinkedIn Pulse article. An answer engine samples that public artifact. A buyer asks the engine which vendor to choose. The engine recommends. The buyer acts.
Links one through three are the dark funnel. That chain is how untracked mentions influence AI answers, and this guide is about who controls it.
The numbers explain the urgency better than any pep talk. Roughly 73% of the B2B buying journey now happens in the anonymous phase before a vendor is contacted. Somewhere between 70% and 80% of the decision is made before a rep is engaged, and 61% of buyers say they would prefer a fully rep-free experience. The average journey runs about 272 days across roughly 88 touchpoints, and almost none of them are yours to log.
Then AI compounds it. Around 93% of AI search sessions end with no click to any cited site, so your analytics see nothing at all. Meanwhile a typical enterprise B2B brand is named in only about 3% of the AI responses relevant to it. That leaves roughly 97% of recommendation moments happening without you in the room.
Pro tip: stop calling this an attribution problem. Attribution was the old framing, when the dark funnel only influenced deals you eventually closed. Today it decides whether you get recommended at all, which makes it a distribution problem you can actually work on.
Step 1: Map the surfaces where AI samples your category
A dark funnel AI search program starts with evidence, not instinct. Your first job is to see the B2B off-domain conversations AI engines actually read when someone asks about your category.
Build a prompt set of 20 to 50 buyer-intent questions. The commercial ones, not the definitional ones: "best X for [use case]," "X vs Y," "alternatives to X." Run every prompt across ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode. Capture every cited URL.
Now tag each URL by domain type: subreddit thread, review-site profile, LinkedIn post or Pulse article, YouTube video, Quora answer, Wikipedia entry, trade editorial, analyst page, newsletter. Add one more column that changes everything: controlled or uncontrolled. Controlled means you can post there. Uncontrolled means you cannot.
Expect 60% to 70% of your citations to sit in the uncontrolled column. That is normal, and it is the whole point.
The distribution across surfaces is lopsided in ways worth knowing before you spend a rupee or a dollar. Reddit holds roughly 3.11% of global AI citation share, the single most-cited domain across the major engines, and 99% of ChatGPT's Reddit citations point at individual discussion threads rather than subreddit indexes. YouTube sits second at about 2.13%, with 85.4% of those citations pointing at a specific video. LinkedIn is only about 0.41% globally, yet it is the most-cited domain for professional queries across six engines. Google's AI surfaces lean hard on YouTube at roughly 23% and Wikipedia at roughly 18%.
The headline for B2B is blunter still: around 85% of citations on B2B category queries come from third-party review sites, directories, and editorial guides, not from brand sites.
You are done with this step when you have a spreadsheet of citations by prompt, by engine, by domain type, with the controlled column filled in.
Common mistake: running this once and calling it tracking. Models resample constantly, and citation shares swing. ChatGPT's Reddit share moved from about 7% to about 1% and back to about 3% inside a single year. Treat the spreadsheet as a living artifact.
One more caution while you read the market's numbers. You will see a widely repeated claim that 40% of ChatGPT citations come from Reddit. That figure conflates Reddit content surfaced through Google's index with direct citation. The lower, transparently measured share is the safer one to plan against.
Step 2: Name your three to seven truth committees
Here is the good news buried in your audit. You are not fighting the whole internet.
Aggregate your citation list by domain and look for repeats. Across your highest-intent prompts, a small set of URLs and surface types will keep showing up. That set is your truth committee: the handful of off-domain sources answer engines default to when someone asks about your category.
Most B2B categories have three to seven. Not thirty.
Write a one-page list. For each committee, note the prompt cluster it serves, which engines lean on it, and roughly how much of the citation share it carries. That page is your investment map for the next two quarters.
You are done when you can point at a specific subreddit, a specific review category, a specific publication, or a specific creator, and say "this is where the recommendation gets made."
Common mistake: writing "win Reddit" on the plan. Reddit is not one place. Different threads in different subreddits answer different prompts, and specificity beats scale every time. The same is true of review sites and LinkedIn.
Step 3: Baseline mention rate, citation rate, and share of voice
Do not touch a single surface until you have a baseline. This is the step people skip, and it is the reason so many AI visibility programs cannot prove anything six months later.
Run your prompt set three times over a two to four week window, across every engine you care about. For each run, log four things:
- Mention rate: how often an engine names your brand at all.
- Citation rate: how often it links to one of your pages as a source.
- Share of voice: your slice of total citations against a named competitor set.
- Visibility trend: the movement between runs.
Stratify by prompt cluster. "Best X" behaves differently from "X vs Y," and both behave differently from "alternatives to X." Averaging them hides exactly the signal you need.
Keep mentions and citations in separate columns. They are not the same thing, and treating them as one number is how teams end up celebrating noise.
You are done when you have a number per cluster, per engine, per competitor, with a date stamp. Every action you take from here gets compared to it.
This is one place DeepSmith does the heavy lifting instead of you. Its AEO module runs your tracked prompts on a schedule and reports mention rate, citation rate, and share of voice with a per-platform breakdown, a competitor leaderboard, and the sources AI cites most for your prompts. The Prompts view keeps per-prompt history so you can see drift rather than guess at it, and Discover Prompts generates a starter set from your product, persona, and buyer-stage context if you are staring at a blank page. Engine coverage rises by plan: Pro tracks ChatGPT, Grow adds Perplexity, Scale adds Gemini, and Enterprise covers all five.
Common mistake: acting first and measuring later. Without a baseline you will never know whether the Reddit answers worked, the reviews worked, or the market simply moved.
Step 4: Seed the surfaces your committees actually live on
Now you act, and you act narrowly. Each committee gets its own play, because what earns a citation on one surface gets ignored on another.
A few principles carry across all of them.
Name entities explicitly. "Attio for venture-backed B2B SaaS sales teams" survives extraction. "A good CRM tool" does not. Engines need the brand, the category, and the segment named together in one clean sentence.
Write things a model could copy. A short top-line answer, then structure, then one honest drawback and one next move. If an engine could not lift a clean fragment out of your post, that post will not earn a citation.
Let named humans do it. Practitioner and executive bylines outperform company pages by a wide margin. LinkedIn company-page status updates score essentially zero citation share, while long-form Pulse articles are the format engines pull from. Only about 2% of LinkedIn URLs get cited at all, so the format you choose matters more than the volume you post.
Ask customers for narratives, not ratings. A five to ten sentence review that walks through the problem, the decision criteria, the integration detail, and a quantified outcome moves AI opinion far more than a star rating with one line under it. Spread those across the review platforms your buyers actually use, and keep the volume balanced.
Design private communities to leak in public. Your Slack and Discord never feed answer engines directly. They matter because they produce public artifacts: a recap thread, a public AMA, a distilled FAQ, a short video answering the question three people asked this week. That is the only route by which dark social AI citations ever happen, and it is the mechanism that makes the community conversations AI recommendations rest on visible at all.
Give it six to twelve weeks before you expect movement on a targeted surface. Sustained share-of-voice gains against an entrenched competitor take longer, often six months or more.
Common mistake: posting brand awareness content with no citation outcome in mind. Warm, on-brand, unquotable content is invisible to an answer engine. So are promotional replies, affiliate links, and new accounts with no history.
And no, you cannot shortcut this with paid seeding. Retrieval favors authentic conversational signals, and obvious placements get downweighted. Genuine participation is slower and it is the only thing that holds.
Step 5: Mirror every off-domain win as an owned page
Off-domain signals get you named. Owned pages get you cited. You want both, and this step is the bridge.
Every time an off-domain artifact performs, a Reddit answer that keeps getting quoted, a Pulse article that travels, a review excerpt that shows up in answers, turn it into a durable page you control. Same claim, same specificity, better structure.
Build the page the way engines read: a definitional answer near the top, a comparison table where a comparison is implied, an FAQ block, schema markup, internal links to the rest of your cluster. Structured tables get pulled far more often than the same information written as prose.
You are done when each high-impact off-domain signal has a matching owned page an engine can cite as the source, and when your own pages start appearing in the citation column of your tracking sheet.
Refresh those pages on a short cycle. Recently updated content earns meaningfully more citations than the same content left to age, and freshness is one of the few levers that works on every engine.
Common mistake: treating an off-domain win as a one-off placement. It is research. It told you exactly which phrasing of your value proposition survived contact with a real buyer, and that phrasing belongs on a page you own.
Step 6: Benchmark competitors surface by surface
You are not measuring yourself in a vacuum. Share of voice is a zero-sum number, and somebody is currently holding the share you want.
For each prompt cluster, list the brands the engines cite today. Capture their share against yours, the exact pages winning the citations, and the surfaces where they dominate. Build a gap map with three axes: domain type, competitor, engine.
Then be selective. Invest where the gap is large and your buyer genuinely lives. Defer where a competitor is strong in a channel your ICP ignores, because winning there produces activity metrics and no pipeline.
This is the second place DeepSmith carries real weight. Its Competitor citations view shows who wins citations for your prompts, on which exact pages, and how each competitor performs by platform, so the gap map builds itself instead of eating a Friday. Content Intelligence tracks what each competitor publishes as it ships, and Remix turns a competitor page that is working into ready-to-use idea titles that land in your Idea Bank.
You are done when you have a living gap map that tells you where to spend next, refreshed on a schedule rather than when someone remembers.
Common mistake: copying a competitor's surface mix wholesale. Their truth committees are not automatically yours, especially if you sell to a different segment.
Step 7: Run the loop weekly, because models keep resampling
If you only change one thing after reading this, make it cadence. The B2B off-domain conversations AI systems sample are refreshed constantly, so influence without repetition decays.
Here is a rhythm that holds for a small team:
- Weekly: one or two off-domain drops on your priority committees, plus a prompt re-run and a citation-share update.
- Monthly: refresh the truth-committee list, since surfaces rise and fall.
- Quarterly: revise the competitor gap map and rebalance where the effort goes.
Good looks like an upward trend across quarters on mention rate, citation rate, and share of voice on at least three of the five engines, with the movement traceable to specific off-domain actions you took.
Bad looks like a one-time audit declared a win, vanity mentions counted as citations, and activity launched without a baseline.
DeepSmith closes this loop on the production side. Ideas from Remix, tracked prompts, and topic gaps become Planned Content, then finished articles through the Writer, with research, internal links, schema, metadata, and a cover image built in. Autowrite generates scheduled pieces on their dates so the cadence survives a busy week, and the Apps Library turns each published article into platform-native versions for LinkedIn, Reddit, X, Medium, Substack, newsletters, and Slack or Discord. Deep IQ keeps your positioning, voice, persona, and content types consistent across all of it.
Common mistake: running this as a project with an end date. It is an operating cadence. The teams that win here are not the ones with the biggest budget, they are the ones who did not stop.
What to do next
Do not try to run all seven steps this month. Take the first two.
Pick 20 prompts your buyers actually ask. Run them across the engines you care about. Write down every cited URL and circle the ones that repeat. That single afternoon will tell you more about how untracked mentions influence AI answers in your category than another quarter of blog posts ever will.
Then pick one truth committee. Just one. Work it for six weeks, keep your baseline honest, and let the trend tell you whether to double down or move on.
If you want the measurement half handled for you, DeepSmith tracks your prompts across ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode, shows which pages and competitors win each citation, and produces the on-brand content that closes the gaps, in one platform. You can start a free trial and see real data and real drafts before you pay.
You are closer to this than you think. Dark funnel AI search is early enough that most of your competitors have not drawn the map either.
