Measurement
We measured 3,538 AI answers. Top-of-funnel content got named zero times.
By Arnav Mukherjee, founder of TofuBofu · August 10, 2026
When we built the scoring for this product we had to pick weights. Every visibility score is a weighted average of something, and the honest version of that decision is that somebody sits down and decides which questions count for more. We put bottom of funnel at 50 percent, middle at 30, top at 20.
That is a strange thing to do if you believe the funnel works the way the diagram says. The diagram says a buyer floats in at the top, gets educated, gets nurtured, and eventually arrives at the bottom ready to compare vendors. Under that model the top is where the volume is and the bottom is a rounding error you optimise last.
We weighted it the other way round because of what the answers look like. A buyer opens ChatGPT and types "best managed IT provider for a 40 person law firm" as their very first move. That is a bottom-of-funnel question. It is also their first question. The engine responds with five named companies, and if you are not one of them, everything you published at the top of the funnel was addressed to a search that buyer never ran.
That was the reasoning at the time and it was a judgement call. Since then we have run enough scans to check it. Across 3,538 question-and-engine cells, the top of the funnel returned a mention rate of zero, and the weights turn out to have been generous to it.
The funnel was never wrong. It just assumed a search box.
TOFU, MOFU and BOFU describe a real sequence. People genuinely do move from "I have a vague problem" to "I am comparing two vendors", and they genuinely do need different content at each point. If you want the full model, channel by channel, we wrote the complete TOFU, MOFU, BOFU guide separately and none of it has stopped being true.
What the model quietly assumed is that a buyer could only ask one small question at a time. A search box takes two to five words and returns ten blue links. It cannot answer "who should I hire", so the buyer had to assemble that answer themselves out of many small searches, and that assembly process is the funnel. Awareness first because awareness was all the search box could deliver first.
An assistant removes that constraint. It will take the whole situation in one go, including the company size, the industry, the trigger event and the constraint, and it will return the shortlist immediately. The stages did not disappear. They collapsed into a single exchange, and the order flipped, because the first thing a person asks a machine that can answer anything is the thing they actually want to know.
The same buyer, two eras, opposite entry points
The number that settles it: zero
Everything above is an argument. Here is the measurement, and it is blunter than the argument.
We went back through every completed scan in our database and pulled 3,538 question-and-engine cells: one cell for each time we put one buying question to one AI engine on behalf of a real company and recorded whether that company was named. Then we grouped them by what kind of question it was.
Read the bottom three rows again. Across 602 cells covering informational, category-awareness and pure awareness questions, the companies we scanned were named zero times. Not rarely. Zero. And it is not one unlucky brand dragging an average down: six different companies contributed to those 602 cells and every one of them scored zero independently.
This is the top of the funnel. It is the content every marketing playbook tells you to start with, the educational article that explains the problem before you sell the solution. In our data it produced no mentions at all, while the buying question produced 5.4 percent and a question that names the company outright produced 56.9 percent.
Two honest caveats. First, companies come to us because they suspect they are invisible, so this is a sample selected on the problem and the absolute levels will be worse than a random B2B company's. What survives that is the shape: the comparison between rows is within the same sample, and the gap between 0 percent and 56.9 percent is not a selection artifact. Second, one company accounts for 336 of the 602 informational cells, so the zero leans on it more than on the other five, all of which were also zero.
Four more findings from the same data
Supporting evidence from our scans and research, because this argument is worth nothing as an opinion.
One engine barely answers the buying question at all
Google AI Overviews answered roughly 9 percent of the commercial vendor queries in our data, against roughly 50 percent of questions about a specific named brand. Google frequently renders no AI Overview on a commercial query, so the slot is simply silent on the exact questions that decide purchases. The conversational engines answer them readily. This is the single strongest argument against reading one blended visibility number: the surface where buying happens is not uniform across engines.
Buyers phrase the buying question as a list request, not a qualified brief
We pulled the phrasings people actually search across 10 real vendor categories and got 90 of them. They are blunt and list-shaped: 'list of MSP companies', 'top 10 best managed it services', 'top MSPs', 'which IT service company is best', 'how much should managed IT services cost'. They averaged 5.5 words. The questions a language model writes from a company's own website averaged 7.8 words and were far more qualified. Neither is wrong, but only one of them has evidence that a human phrased it that way.
On local buying questions the engines barely agree with each other
Across five US metros we collected 136 firm-and-market entries from four answering engines. Eighty five percent were named by exactly one engine and not a single firm appeared on all four. On broader category-level questions asked of the same engines on the same day, two firms were named unanimously. Agreement is a function of how specific the question is, and buying questions are specific, so at the bottom of the funnel you are running four separate races.
Being listed in the obvious directory is not the same as being named
In that study, the engines that answered were asked who the best managed IT providers were across 15 regions, and every answer was checked against each region's own Clutch directory. The directories held 228 firms between them. Engines named 89 of those and skipped the remaining 139 entirely. Presence in the place a buyer would once have checked by hand does not transfer automatically into the answer a buyer now gets instead.
Put those together and the practical instruction is unambiguous. If you are absent from the answers to your category's buying questions, that is the emergency, and it is not fixed by publishing more awareness content. It is fixed by becoming the kind of company an engine can confidently place on a list.
BOFU: nine plays, in the order I would run them
Start here even though it feels backwards. These are ordered by leverage per hour, not by difficulty.
Write down the ten questions that decide your deals, in your buyer's words
Not your words. Take them from sales call recordings, inbound emails and the People Also Ask block for your category. If your list contains the phrase 'end-to-end' or your own product category name as you coined it, you have written marketing copy rather than buyer questions. The test: would a stranger type this into a chat window at 11pm.
Ask those ten questions to every engine yourself and write down who gets named
This takes an hour and it is the only way to know whether you have a visibility problem or a positioning problem. Read the full answers, not just whether your name appears. The companies named beside you are the comparison set your buyer is being handed, and they are frequently not the competitors you think you have.
Publish a page that answers 'how much does this cost'
Cost is the question buyers ask that firms most consistently refuse to answer. When we categorised the descriptive language in 64 real AI answers recommending vendors, price language appeared in 38 percent of them, the lowest of any attribute we counted, precisely because it is the one thing firms deliberately omit. If your site says nothing about cost, the engine describes you without it, or worse, infers it.
Build one comparison page per competitor the engines actually name beside you
Not per competitor on your internal battlecard. The set the engines use is the set your buyer sees. Make the page honest, including where the other option is genuinely better, because a comparison page that never concedes anything reads as marketing to a model as well as to a human.
Get onto the list-shaped third-party pages
The engines answering 'top 10 X' reach for pages that already contain lists of named vendors. Those are directories, industry rankings, roundups and review sites, and almost none of them are your domain. Breadth of independent presence is the lever with an actual handle on it.
Treat reviews as a ranking input, not a reputation nicety
Kevin Indig's analysis with G2 found G2 holds about 22.4 percent share of voice for software, and that roughly 10 percent more reviews associates with about 2 percent more AI citations. That is a slow lever but a real and compounding one, and unlike most of this list it also works while you sleep.
Be present where buyers go for unpaid opinion
When we pulled observed demand for managed IT, 'best managed it services reddit' came back twice without anything in our seeds pointing there. Buyers append that word because they want an opinion nobody bought. Participate honestly under your own name; do not manufacture threads, which is both detectable and against every platform's rules.
Put your geography and your buyer's size in plain readable text
In those same 64 answers, geography appeared in 88 percent of the descriptions engines wrote about the firms they named, and size-fit language in 62 percent. If the only place your service area exists is a map embed, and the only place 'we work with 20 to 200 person firms' exists is a sales deck, the engine will describe you without either and you will lose the qualified questions.
Add the structured data, then stop thinking about it
SE Ranking found 71 percent of ChatGPT-cited pages use structured data, and 65 percent for Google AI Mode. Organization, Service, FAQPage and Offer where they genuinely apply. This is a floor, it is a day of work, and it is not a strategy. Anyone selling you schema as the answer to AI visibility is selling you the cheapest part of the job.
Find out which buying questions you are missing from
A free scan runs your category's buying questions across all six engines and shows you every answer in full, including who got named instead.
Get your free auditMOFU: six plays for the questions that qualify you out
Middle of funnel used to mean nurturing. In an AI answer it means the qualifying clauses the engine attaches to your name, which decide whether the buyer keeps you on the list after the first response.
Publish one page per vertical you genuinely serve
Vertical language showed up in 72 percent of the descriptions engines wrote about vendors in our sample. 'IT services for law firms' is a different entity in a model's understanding than 'IT services', and only one of them survives a buyer who says 'but we are a law firm'.
State your compliance and credentials as text on a page
Compliance language appeared in 73 percent of those descriptions. SOC 2, HIPAA, ISO, the certifications your engineers hold. A badge image with no accompanying text is invisible to the thing writing the answer.
Answer the objection you lose on, in public
Every firm has one recurring reason it loses. Write the page that addresses it directly and honestly. This is the content most likely to survive a buyer pushing back inside the conversation, which is where the shortlist actually gets cut.
Build a real FAQ, then mark it up
Not six questions invented by a copywriter. The actual questions your sales team answers twice a week, in the phrasing buyers use. FAQ markup is the highest-value single piece of structured data for this channel because the format matches how the answer is assembled.
Name your tenure and your size honestly
Tenure language appeared in 56 percent of engine descriptions. 'Founded 2013, 34 people' is a fact an engine can repeat. 'Established leader' is not.
Make your case studies findable and specific
A case study behind a form does not exist as far as an engine is concerned. One that names the industry, the size of the client and a concrete outcome gives the model exactly the clause it needs to justify recommending you to someone who matches.
TOFU: five plays, and a warning about sequencing
Top of funnel content has not stopped mattering. Its job changed. It used to earn a reader directly. Now its main job is to be evidence: material an engine can read that establishes what category you belong to and what you are credible at, which is what lets it name you confidently when somebody else asks the buying question.
Publish the explainer for your category, once, properly
One genuinely good page on what your category is and how to buy it does more than twenty thin posts. It is the page that teaches a model what words belong next to your name.
Say the same thing about yourself everywhere
Your site, your directory profiles, your review-site descriptions and your social bios should describe your category, your geography and your buyer in consistent language. Inconsistency across sources is what stops an engine being confident enough to name you, and it costs nothing but an afternoon to fix.
Put original data into the world
You can out-publish nobody on general advice, and everybody on facts only you hold. Numbers from your own operations, your own client base or your own experiments are the most citable asset a small firm can manufacture, because they cannot be sourced anywhere else.
Use video where your category is genuinely visual
YouTube is a distinct citation surface and is treated as a heavyweight source by some engines. Worth it if you have something to show, not worth it as a checkbox.
Do not start here if you are invisible at the bottom
This is the sequencing warning. Awareness content compounds slowly and pays off through corroboration, which is a second-order effect. If your category's buying questions currently return five competitors and not you, spending this quarter on thought leadership is choosing the slow lever while the fast one sits untouched.
Four things that look like plays and are not
Rewriting your homepage again
The engine is not reading your homepage harder than it did last quarter. It is reading everyone else's pages about you. We enriched 72 IT firms our research had already measured as absent from AI answers, expecting under-marketed companies, and found 27 running full marketing automation and 28 running their site on HubSpot CMS. The stack was not the problem and neither was the copy.
Buying domain authority
Nofollow links from high-authority domains, directory spam and link packages are answering a question nobody is asking here. An engine deciding who to name is looking for independent sources that agree about you, not for a number in an SEO tool.
Manufacturing reviews or Reddit threads
We were asked about fabricated review profiles and declined. They violate every platform's terms, they are illegal to present as genuine, they get wiped, and the cleanup costs more than the work you avoided. Everything in play 6 and play 7 works only if it is real.
Optimising for one engine and calling it done
Eighty five percent of the local firms in our five-metro study were named by exactly one engine of four. If you check ChatGPT, see yourself, and stop, you have measured one of four races and assumed you won the other three.
Where this leaves the funnel
Not dead, and not even much changed as a description of what buyers need. What changed is that the bottom of it became the front door, and the two stages above it turned into evidence the machine reads on the buyer's behalf rather than journeys the buyer walks personally.
That reordering is the whole reason this company is called what it is. Tofu and Bofu, the two ends of the thing, with the argument that the end everyone treated as last is now the end you have to win first.
G2's 2026 buyer research found 51 percent of B2B buyers now begin vendor research on an AI chatbot, up from 29 percent, that 69 percent had changed their vendor choice based on what an AI told them, and that one in three ended up with a vendor they had not heard of before the AI named it. Read that last number again with the funnel diagram in mind. A third of these deals begin with the buyer at the bottom of the funnel, being introduced to a company they had zero prior awareness of. The awareness stage did not precede the decision. It was skipped, and the decision manufactured it.
Which is also why the old line still holds: SEO is the floor, and AEO is a distinct layer you can win even when your Google rank does not move. Being crawlable and cleanly structured is necessary and cheap. Being the company four independent sources agree is the right answer to a specific buying question is the part that wins, and you can start on it this week regardless of where you rank today.
Frequently asked questions
What do TOFU, MOFU and BOFU mean?
Top, middle and bottom of funnel. TOFU is awareness content for people who do not yet know they have a problem, MOFU is consideration content for people comparing approaches, and BOFU is decision content for people choosing a vendor. The model is decades old and it describes a real sequence of buyer questions. What has changed is not the stages but the order a buyer reaches them in, because an AI assistant will hand out a vendor shortlist in response to the very first thing a buyer types.
Why is BOFU more important than TOFU for AI search?
Because the bottom-of-funnel question is now often the first question. A buyer who types 'best managed IT provider for a 40 person law firm' has skipped awareness and consideration and gone straight to asking for the shortlist, and the engine answers with named vendors. If you are not in that answer you are not in the deal, and no amount of awareness content further up compensates, because the buyer never ran the awareness search. We weight our own scoring 50 percent BOFU, 30 percent MOFU, 20 percent TOFU for exactly this reason.
Do AI engines answer bottom-of-funnel questions?
Most do, and one largely does not, which is worth knowing before you read any blended score. In our own measurement Google AI Overviews answered only about 9 percent of commercial vendor queries while answering around 50 percent of questions about a named brand, because Google frequently renders no AI Overview at all on a commercial query. The conversational engines answer vendor questions readily. So the surface where the buying decision happens is not the same surface across all six engines, and a single averaged number hides that.
What kind of BOFU content gets you cited by AI?
List-shaped, third-party and specific, in that order of leverage. When we pulled the phrasings people actually search in vendor categories, the demand was blunt and list-shaped: 'list of MSP companies', 'top 10 best managed it services', 'how much should managed IT services cost'. Engines answering those questions reach for pages that already contain lists of named vendors, which are usually directories, roundups and comparison pages on domains that are not yours. Your own site's job is to make you easy to place on those lists and easy to describe correctly once you are.
Does top of funnel content still matter for AI visibility?
Yes, but not for the reason it used to, and our own data is harsh about the old reason. Across 602 question-and-engine cells covering informational, category-awareness and pure awareness questions, the companies we scanned were named zero times, against 5.4 percent on the buying question. TOFU content no longer earns the mention directly. Its job now is corroboration: material an engine can read that establishes what category you are in, who you serve and what you are credible at, which is what makes an engine confident enough to name you when somebody else asks the buying question. That is a real function and worth funding. It is just not the place to start if you are absent from the answers where money changes hands.
How do I find the BOFU questions my buyers actually ask AI?
Take them from four sources and stop guessing. First, your own sales calls and inbound emails, which carry the phrasing buyers used before they found you. Second, Google's People Also Ask and related searches for your category, which record how humans phrase the need even though the register is shorter than a real prompt. Third, your Search Console, where AI Mode follow-up turns now appear as separate queries. Fourth, ask the engines the questions yourself and read who they name. When we compared questions written by a model from a company's own website against observed search demand across 10 real scans, the overlap was zero out of 90, so the generated set alone is not enough.
Should I build one page per funnel stage?
Build one page per buying decision, which is not the same thing and is usually fewer pages. A stage is a category of intent, a decision is a specific question a buyer needs answered before they can move: what does this cost, who is this for, how does it compare to the obvious alternative, is it safe to buy. Splitting near-identical intent across separate thin URLs is the most common self-inflicted wound here. One strong page that fully answers a decision outperforms three weak ones aimed at three phrasings of the same decision.
Sources and further reading
- TofuBofu, AI visibility for managed IT providers, 2026: the 15-region study behind the 228 directory-listed firms and the 139 named by no engine, with full method.
- TofuBofu, five-metro engine disagreement study: the 136 firm-and-market entries and the 85 percent named by exactly one engine.
- TofuBofu, what engines say about the vendors they name: the categorised descriptive language across 64 real answers, including price at 38 percent, geography 88 and compliance 73.
- G2 2026 B2B buyer research: 51 percent begin vendor research on an AI chatbot, up from 29 percent, 69 percent switched vendor based on AI, and one in three chose a vendor unknown to them before the AI named it.
- SE Ranking, 2026: 71 percent of ChatGPT-cited pages use structured data, 65 percent for Google AI Mode.
- Kevin Indig with G2: G2 holds about 22.4 percent share of voice for software, and roughly 10 percent more reviews associates with about 2 percent more AI citations.