Talkadot grew its AI visibility 2.2× in 9 weeks.

The situation

Talkadot is a platform that helps event planners find and book professional speakers using real audience feedback data, and helps speakers capture feedback, testimonials and leads through a simple QR code.

It is a two-sided business. Speakers use the product. Planners are where the revenue fires, because a booking is how Talkadot gets paid.

More and more of those planners start by asking an AI. How do I find a keynote speaker? How much does one cost? How do I know they are any good? The answer the AI writes back is the new shortlist.

Talkadot had something nobody else in the category had: more than a million verified audience survey responses, collected across tens of thousands of real talks. It also had a cofounder and CEO, Arel Moodie, who had already written the sharpest planner-side articles in the space.

None of that was showing up where planners were asking.

The April baseline made it plain. Talkadot scored 44 out of 90 on our 9-dimension AI visibility scorecard, a D. Of the 20 searches the cofounder had named as the ones that mattered, Talkadot appeared in 4. Three of those 4 had the word "Talkadot" in them.

AI knew Talkadot when you asked for it by name. It didn't think of Talkadot when a planner asked how to find a speaker.

If you have the best data in your category and AI only names you when someone types your name, you do not have a content problem. You have a translation problem.

What was in the way

Talkadot was not short on knowledge. It was short on knowledge in a form a machine could retrieve.

• Data locked in a report

  A million-plus survey responses lived in one long report. AI engines cite specific, checkable findings. A long report is not one.

• No pages for planner questions

  Content architecture scored 1 out of 10, with roughly 3 of the 14 page types AI engines lean on. No cost page. No comparison pages. No glossary. No answer to "what is a speaker marketplace?"

• A brand described ten ways

  There was no locked description of what Talkadot is. Every bio, directory listing and article phrased it differently, so AI had nothing consistent to learn. Brand fingerprint scored 0 out of 10.

• Zero organic discovery

  Organic discovery scored 0 out of 10. Google's AI Overviews had never cited the site once.

• Two audiences, one content plan

  Speakers are the users. Planners fire the revenue. A piece written for both lands with neither, and speaker-side search traffic does not book speakers.

The strategy: four decisions

1. Mine before you write

   Before making anything new, we inventoried what Talkadot already knew: one strategy doc, nine articles the cofounder had written, and one industry report. Those became the brand brain, the voice guide, the query bank and the data layer. Net-new writing came after.

2. Use the founder's scoreboard, word for word

   The cofounder's strategy doc named 20 searches and a target of showing up in 8. We tracked those 20 verbatim and added 11 supporting searches for a 31-search bank. You do not paraphrase the success metric.

3. Planner first, 95/5

   If a page would be equally useful to a speaker and a planner, it gets rewritten for the planner. We also cut every "best motivational speakers for X" query. That traffic is researching speakers. It is not booking them through Talkadot.

4. Original data or no page

   AI engines cite original data and skip aggregators. Every page cites at least one finding from Talkadot's own dataset, or is clearly framed as the cofounder's opinion. No generic listicles.

How we did it

1. Capture what the company already knows · Foundation

   First we put a frozen baseline on record: the 31-search bank, a citation scan of 310 AI results, the 44-out-of-90 scorecard, and GA4 and Search Console baselines. Nothing shipped until the starting point was written down.

   Then we worked from the inputs. One strategy doc and nine articles became a brand brain, an ICP profile, a locked one-sentence entity description, a 95/5 planner-first rule, five content pillars with fixed weights (40/25/20/10/5), and a voice guide measured from the cofounder's own writing.

   The voice guide captures his habits: one-line paragraphs, "It is not X. It is Y." reframes, zero em dashes, and a 30-plus-word banned list. It came out of his sentences, not out of a prompt. That is why drafts sounded like him on the first pass.

2. Turn one report into a citation engine · Foundation

   We broke the State of the Speaking Industry 2026 report into 30 data atoms. Each atom holds one finding, its raw numbers, the section it came from, and one approved sentence a page can use word for word.

   Later, when the dataset looked tapped out, we diffed the report's sections against the atoms file. Four sections had never been atomized. That gave us 8 more citable findings without asking the client for any new data. The library now holds 39.

3. Map every search to a page type · Launch, June 16

   Each of the 31 searches was matched to one of 14 AEO page types: what-is, cost, buyer guide, comparison, alternatives, best tools, glossary, FAQ hub, statistics and more. The first batch went live on June 16. Every page follows the same build:

   • a bolded, liftable answer in the first 60 words
   • one section per buyer question
   • 3 to 5 FAQs written as standalone answers
   • structured data (schema) so a model can parse it block by block
4. Ship through gates, not reviews · Weeks 1–9

   Every draft clears three gates before the cofounder sees it.

   • Fact gate. Every number traces to a data atom by ID, and every derived figure is recomputed from the raw numbers.
   • Voice gate. Banned words and cadence rules fail the draft automatically.
   • Claim gate. An independent reviewer running on a different AI model from the writer reads the page against the brand rules. That reviewer is the only publish authority.

   The gates earned their keep. They caught four errors that would have shipped:

   • a third-party statistic that did not exist
   • a multiplier in our own data file that didn't match its raw numbers (6% against 1% is 6×, not "nearly 5×")
   • a fee figure invented by joining two real numbers from different cuts of the dataset
   • a comparison page that told only one side of a pricing story

   A human skim passes all four. The cofounder approves drafts. He does not rewrite them. By week 9, 31 pages were live: 14 of 14 page types, more than 80,000 words, and every page passed the cross-model gate.

5. Publish into the client's own stack

   Pages publish straight into Talkadot's Webflow through scripts that re-check the live page after every write. We learned that one the hard way. Webflow returned a success code on a write that silently did nothing, and 8 feature images never reached the live site until we added the check.

   Feature images are resized and converted before upload. Across 31 images, the site went from 101MB of images to under 1MB.

6. Compound

   Every correction becomes a dated, permanent rule the system checks against. Some examples are the em dash, the right award attribution in a bio, and how to frame pricing honestly. There are 51 rules now.

   When the original 31 searches saturated, we re-derived the bank from the live answers, which turned up 38 new searches and 7 open gaps to work next. The system that wrote page 31 is better than the one that wrote page 1, because it cannot repeat a mistake it has already made.

The results

The same 31 searches, the same script and the top 10 cited sources per answer, measured on Perplexity at the April 23 baseline and again on August 17, 9 weeks after launch:

Tracked AI answers featuring Talkadot, out of 31 · 2.2×

Before

13

After

28

Priority searches won, out of 20

Before

4

Goal

8

After

19

Perplexity citation scan, fixed 31-search bank, top 10 cited sources per answer. Baseline April 23; latest August 17.

Traffic and signups come from GA4 AI referral sources, normalized per day: a 28-day window before the engagement against the first 60 days after launch.

Three more numbers from the August scan show how far it went:

• 21 of 24 planner searches without "Talkadot" in them now feature Talkadot.
• 23 of 31 tracked searches cite Talkadot as the #1 source.
• 97 citations to talkadot.com across the 31 answers, about 3 per answer.

From brand answers to category answers

The number that matters most is not 19. It is which searches moved.

In April, Talkadot mostly showed up when someone typed its name. By week 9, it was in the answer for the questions planners ask before they know any vendor's name:

• How do I find a keynote speaker for a corporate event?
• How much does a keynote speaker cost?
• What is a speaker marketplace?
• How to vet a professional speaker
• Speaker bureau alternatives

That is the difference between being found and being chosen.

We had a goal of winning 8 out of the 20 priority searches for our particular tool, and we ended up winning 19 out of 20.

What changed for the business

• AI traffic that converts

  ChatGPT referrals grew 3.5×, and AI referrals produced the first 5 signups ever sourced from AI search. As the cofounder put it, it wasn't just vanity metrics.

• Category answers, not brand answers

  Planners who have never heard of Talkadot now meet it in the answer to "how do I find a keynote speaker?" That is where the shortlist gets made.

• A founder who approves

  Drafts reach the cofounder already checked against his data, his voice and his banned-word list. His job is judgment, not assembly.

It wasn't just vanity metrics. We roughly saw three times the referral traffic coming from AI, and we've actually had new signups for Talkadot because of being discovered through AI searches.

Steal this: five moves you can run this week

You don't need us to start. Here is the first week of the Talkadot install, in the order we ran it.

1. Write down your 20 searches

   List the 20 questions your buyer asks before they know your name. Skip anything with your brand in it. Put each one into Perplexity and ChatGPT, and count how many answers name you. That count is your baseline. Write it down before you change anything.

2. Lock one sentence

   Write a single sentence that says what your company is, who it serves and what makes it different. Then use it everywhere: your homepage, LinkedIn, G2, Crunchbase, every bio. AI learns you from repetition. Talkadot scored 0 out of 10 here in April.

3. Atomize your best data

   Take your most original data source, whether that's a report, a survey or product usage. Break it into one-finding-per-entry atoms, each with its raw numbers and one sentence a page can quote. Talkadot turned one report into 39 citable findings.

4. Map each search to a page type

   For each of your 20 searches, decide which page should answer it: a what-is page, a cost page, a comparison, alternatives, a buyer guide or an FAQ. Put the answer in the first 60 words. Then build the missing pages first.

5. Review on a different model

   Before anything publishes, have a different AI model from the one that drafted it check every number and claim against your source files. Same-model review misses what the writer talked itself into. That gate caught four errors on Talkadot that a human skim would have passed.

Moves 1 and 2 take an afternoon. Moves 3 through 5 are where most teams stall. That's the part we install.

What got installed: the system is the asset