The DemandBox blog
The guides are the neutral reference. This is where we take a position. Short posts on what is actually working in AI search and demand generation, what is theatre, and what we would do on Monday morning.
Written by Avishai Sam Bitton, Founder, DemandBox.
Subscribe via RSSA defensible AI search competitive analysis needs a fixed prompt set, a repeatable sampling method, and a record of what was cited, not a one-off screenshot of ChatGPT's answer.
Start here
Good. That is usually the most useful call we have all week.
A GEO agency and an AEO platform solve different problems, and buying the wrong one first wastes a budget cycle you will not get back.
Google gives you no direct report on AI Overviews citations, so tracking visibility there means building your own sampling process and accepting its limits.
Buy the parts of creative production that require volume and speed, keep the parts that require deep product and buyer knowledge in house, and judge every piece of output against a real competitor's ad, not against a mood board.
Absence from AI answers usually comes from a small, ordered set of causes, and most teams start fixing the hardest one first because it feels like the most important, not because it is the most fixable.
The pitch decks from demand generation companies look interchangeable, so the only comparison that tells you anything is scope of work, ownership of the numbers, and what happens on the way out.
Answer box optimization is a formatting discipline, not a keyword trick: the page has to answer one question completely in the first block a machine can extract cleanly.
Demand capture and demand creation solve different problems on different timelines, and most budget fights happen because a company is funding one while measuring it against the other.
Most flat pipeline problems get fixed in the wrong order, with new budget thrown at channels before anyone checks whether the definitions and routing underneath are lying.
Demand generation marketing covers the programs that build category awareness and buyer trust before a purchase intent exists, and most of what gets sold under that label is lead capture with a better name.
GA4 can track ChatGPT referral traffic with the right channel group, but the number it reports is a floor, not the real total.
Demand generation is the work that happens before a lead exists: building awareness, trust, and a category preference so that when a buying need shows up, your name is already on the shortlist.
The best demand generation agency for a B2B company in 2026 is not a name, it is the model that fits that company's stage, spend and internal capacity, judged against a fixed scorecard.
Brand visibility in AI search improves in a fixed order, and publishing more pages is the last step, not the first.
When you buy an AI visibility tool you are buying a sampling method, and the score it prints is only as meaningful as that method.
Signal based selling fails at routing and follow up, so buying more signal sources makes the problem worse rather than better.
Most AI visibility tracking is a single prompt run on a single day reported as a trend, and it will get you fired for reading noise as signal.
A demo request already told you everything a scoring model is trying to guess. Qualification should decide how fast to route it, not whether it deserves the points.
A growth partner's first 90 days should be written as a contract with four hard checkpoints, not a list of deliverables, and the only checkpoint that counts on day 90 is a pipeline number you can trace back to a source.
LTV:CAC is a ratio you can make say almost anything by picking a churn assumption. CAC payback is a fact, built from cash already collected, and it is the one efficiency number a board will actually push back on.
A single percentage-of-revenue rule cannot govern a marketing budget across three stages that need almost opposite spend, because the job the budget is doing changes faster than the rule does.
Most b2b retargeting spend buys credit for deals that were already closing, and the reported ROAS is arithmetic laundering, not incremental pipeline.
You are not choosing between in house and agency, you are choosing which two or three specific people do the work, and almost no hiring or selection process is designed to evaluate them.
Report volume is a lagging indicator of results. When there is pipeline to show, the deck gets shorter.
Nobody misses a pipeline number because the spreadsheet multiplied incorrectly. They miss because a single optimistic conversion rate was entered eleven months earlier and never revisited.
Attribution models allocate credit according to rules you chose. That makes them a budgeting convention, and treating them as evidence is where the damage starts.
Reddit is where your buyers say what they actually think, which makes it both the cheapest research asset in B2B and a channel that punishes standard marketing instantly.
Audience controls have been quietly deprecated in favour of the algorithm, which means the ad itself is now the primary targeting instrument you control.
In most underperforming paid accounts, the fastest performance gain comes from removing spend rather than optimising it.
An assistant describes you using the consensus of everything written about you, not your homepage, so vague or inconsistent positioning is no longer a brand risk. It is a retrieval failure with a fix list.
Answer engines reward being verifiable and being agreed with, which is a different job from ranking, and almost none of the published advice reflects that.
Sessions were always a proxy for being found. The proxy broke, the underlying thing did not, and most teams are panicking about the wrong number.
The MQL exists to settle an internal dispute about effort. It was never designed to predict revenue, and it does not.
The default agency contract pays for motion, not movement, and every incentive downstream of that contract follows the money.
The guides cover the same ground in depth, with the methods, benchmarks and sources behind these arguments.
Browse the guides
Talk to an Expert