Free resource · 16 Nov 2026

Free playbook · 8 agents

Gemini 4 AI Search Team

Eight agents that measure, fix and report your AI search visibility, set up in one afternoon on Gemini 4 Hydrogen.

Will Leatherman

Will Leatherman

Founder, Catalyst

8
Agents
8
Paste blocks
40 min
Setup time
25 min
Read
See the agents
Will Leatherman, Founder, Catalyst
Founder, Catalyst

I'm Will Leatherman, founder of Catalyst.

Catalyst is a content-led GTM agency. We get B2B companies cited when buyers ask ChatGPT, Claude, Gemini and Perplexity who to buy from.

Free audit

Before you build any of this, find out where you actually stand.

Our free AEO audit scans ChatGPT, Claude, Gemini and Perplexity and shows you what percentage of buyer-intent queries in your category mention you, and what that gap is costing in pipeline. 60 seconds, no account.

Enter your email for a free search analysis
01 · Read this firstFree resource · 16 Nov 2026

01

Read this first

This gives you eight agents that check whether ChatGPT, Claude, Gemini and Perplexity name your company when buyers ask who to buy from, work out why when they do not, and write the fix list.

You end up with one spreadsheet your team reads every Monday and a short list of pages to change each month.

You do not need to code. You need one person who is comfortable following setup steps, which can be you, an engineer or a technical marketer. Setup takes about 40 minutes and happens once.

Start with agent 1 alone. Run it once, read what it wrote, and add the next agent only when the first one is right.

We can run this team for you.

We run AI search for B2B companies. That covers the research, the pages and the placements that get you named in AI answers.

See if we're a fit.
02 · How it worksFree resource · 16 Nov 2026

02

How it works, in plain English

You write 20 questions your buyers would type into an AI assistant, and one person runs those questions through all four assistants and saves every raw answer.

  1. Agent 1 reads that pile of answers and turns it into rows showing where you were named, where a competitor was named instead, and which pages each assistant cited.
  2. Agents 2 to 6 take those rows apart and tell you which pages are feeding your competitors, whether AI crawlers can read your site, which of your own pages have gone stale, and which outside pages you are missing from.
  3. Agent 7 drafts the comparison pages that answer the questions you keep losing, and leaves a marker wherever it needs a number only you have.
  4. Agent 8 pulls your AI referral traffic into the weekly report your leadership keeps asking for.

The eight agents are described below in plain terms, and every block, setting and command sits in one setup section at the end, written for the person who installs it.

03 · What changed with Gemini 4 HydrogenFree resource · 16 Nov 2026

03

What changed with Gemini 4 Hydrogen

Google released Gemini 4 Hydrogen on 12 November 2026. The line that matters for this team is that Hydrogen "pairs native 1 million-token input with up to 1 million output tokens to synthesize dense PDFs, spreadsheets, charts, audio, and recorded video into polished, cited reports and structured deliverables." A month of raw assistant answers, a sitemap export and a GA4 export go into one pass together, and the price is flat across the whole 1 million-token window at $1.50 per million input tokens and $7.50 per million output tokens, so a long input costs the same rate as a short one. You also pick the reasoning depth per job across four thinking levels, low, medium, high as the default, and xhigh. At its default high setting it runs 2.5x faster than Gemini 3.8 Flash, and it cuts fabricated claims by more than 5x. It pulls context from Gmail, Google Drive and Connected Apps like Airtable and Dropbox, and it speeds up Gemini in Google Sheets outputs by over 30%.

04 · Every agent shares one trackerFree resource · 16 Nov 2026

04

Every agent shares one tracker

Everything the agents find goes into one shared spreadsheet with 9 tabs.

AI Answer TrackerShared
TabHoldsWritten by
QuestionsThe 20 buyer questions, the 5 you mark priority, your competitor list and your domainYou
RunsOne row per question, per assistant, per passAgent 1
GapsLists the page behind every answer where a competitor was named and you were notAgent 2
CrawlRecords whether each AI crawler can fetch and read your pagesAgent 3
PagesRanks your own cited pages by age with the one thing to change on eachAgent 4
ThreadsCollects the forum threads where buyers ask what to buy in your categoryAgent 5
ListsScores the roundups and directories the assistants already citeAgent 6
DraftsLogs every comparison page draft with its statusAgent 7
TrafficWeekly AI referral sessions and conversions by sourceAgent 8

No agent writes to another agent's tab. Each one reads whatever it needs from the others and fills only its own, so a bad run damages one tab and you fix it by rerunning one agent.

05 · The 8 agentsFree resource · 16 Nov 2026

05

Each of the 8 agents does one job

Open an agent to see what it does, what you get from it and when it runs.

01Answer SweepTurns the raw answers from all four assistants into one row per question, per assistant, per pass.

What you get. A dated block of rows for the Runs tab, plus seven lines telling you how many questions named you, what you won since last time and what you lost.

Why it works. Assistants give different answers to the same question at different moments, so a single check is a guess. Two passes per assistant and a flag when they disagree is the smallest honest measurement.

Schedule. Weekly on Monday for your 5 priority questions, and the full 20 on the first Monday of the month. Weekly catches movement, monthly keeps the manual question-running under an hour.

Watch out. On run one it will try to answer a question itself and record that as an assistant's answer. Check five rows against the saved answer text. If a row claims you were named and the quoted sentence is missing, the agent filled a gap from memory.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 1.

02Source TraceFinds the exact page that got a competitor named in an answer where you were not.

What you get. A ranked list of the URLs doing the damage, with the page type, whether you appear on it at all, and the three to go after first.

Why it works. Most of the time an assistant names a competitor because one outside page said so. Change what that page says, or get onto it, and you change the answer.

Schedule. Monday, straight after Agent 1. It reads the rows Agent 1 just produced, so it has nothing to work with before then.

Watch out. If everything comes back classed as "other", your category does not fit the page types in the block. Add the type you keep seeing and rerun.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 2.

03Crawl CheckReports whether AI crawlers can fetch your pages and read the words on them.

What you get. One table sorted worst first, with the specific fix for each failed row and the team that owns it.

Why it works. A blocked crawler or a page that only renders its copy after scripts run cannot be cited, no matter how good the page is. This check runs before anything else is worth doing.

Schedule. Monthly, and by hand after every site release or CDN change. Blocks almost always arrive with a deploy, and nobody announces them.

Watch out. If it reports every crawler as allowed and you know your robots file has rules in it, you saved the wrong file. Open what you dropped in the folder and check the first line.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 3.

04Refresh QueuePicks the 10 pages on your site worth updating this month, in order.

What you get. Ten rows naming each page, the date it was last cited, and the one specific thing on it that is out of date.

Why it works. A page an assistant has already cited has proved it can get cited. Bringing that page back into date beats publishing a new one that has proved nothing.

Schedule. Monthly, after Agent 3. Ten updates a month is roughly what a two-person content team ships without stopping everything else.

Watch out. The first run will be thin, because it can only see pages cited in the runs you have already collected. Expect the list to be useful from about week six, and ignore it before then.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 4.

05Thread FinderFinds the forum threads where buyers ask what to buy and checks whether you are in the replies.

What you get. A verified list of live threads, who is being recommended in each, and the three where competitors are named and you are absent.

Why it works. Assistants cite forum threads when buyers ask for alternatives, and those threads stay cited for months. A useful reply in the right thread outlives a blog post.

Schedule. Every two weeks. Weekly is more collection work than the findings justify, and monthly means you arrive after the thread has gone quiet.

Watch out. If it returns a thread you cannot open, it invented it. Click every URL on the first run and delete any row you cannot reach.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 5.

06List ScorerScores the roundups and directories the assistants cite in your category and flags the ones you are missing from.

What you get. Fifteen ranked pages, the route in for each, and the specific reason you belong on it.

Why it works. Best-of lists and directories get cited again and again across different questions, so one correction can move several answers at once.

Schedule. Monthly. Lists update slowly, and pitching the same editor more often than that is how you get ignored.

Watch out. If most rows come back with no route in, the agent stopped at the first screen of the page you saved. Save the footer and the about page too, then rerun.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 6.

07Comparison DrafterDrafts your "vs" and "alternatives" pages in the format of two pages you already like.

What you get. A full draft with a comparison table, a section on when the competitor is the better pick, and a marker wherever it needs a number only you have.

Why it works. Comparison questions are the ones buyers ask right before they shortlist, and the assistants cannot answer them from memory. They have to cite somebody, and it may as well be you.

Schedule. On demand. Run it when Agent 2 shows a competitor winning a question you keep losing, and whenever a new competitor shows up in the Runs rows.

Watch out. Read the "when the competitor is the better choice" section first. If it is vague, or it quietly swings back into a pitch for you, the page loses the honesty that makes it worth citing.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 7.

08Referral ReportWrites the weekly report on sessions and conversions arriving from AI assistants.

What you get. A short report in the format your leadership already reads, with this week against last week and against eight weeks ago, plus the top landing pages.

Why it works. Visibility only matters if it turns into visits and pipeline. This is the number that keeps the work funded.

Schedule. Monday, after Agent 1. Sunday's analytics data is usually complete by then.

Watch out. If the total looks far lower than you expect, traffic from assistants is probably landing in direct. That is a referrer problem on your site, not an agent problem, and the agent should say "not tracked" rather than estimate.

Setting it up. The block to paste for this agent is in the setup section at the end, under block 8.

06 · Setup takes about 40 minutesFree resource · 16 Nov 2026

06

Setup takes about 40 minutes

This section is for the person who installs the team. It happens once. Everything above this point needs no setup knowledge.

  1. 01

    Turn the agent on

    Go to gemini.google.com and sign in with the account that holds AI Pro or Ultra. Open the model picker above the prompt box and choose Gemini 4 Hydrogen. Hydrogen is rolling out across Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search and Google Antigravity, so if the picker does not show it yet, use Google AI Studio with the same blocks.

  2. 02

    Connect the places your files live

    Open Settings, then Apps, and turn on Google Drive and Gmail. Turn on the Connected Apps your team keeps work in, Airtable and Dropbox included. Hydrogen pulls context from all of these, which is how the agents read exports without anyone writing an integration.

  3. 03

    Build the tracker

    Create a Google Sheet called AI Search Tracker with 9 tabs named exactly Questions, Runs, Gaps, Crawl, Pages, Threads, Lists, Drafts, Traffic. Fill in Questions by hand with 20 questions a buyer would type, 10 best-of, 6 alternatives, 4 head to head, none containing your brand name. Mark the 5 with the highest buying intent PRIORITY. Add a column for competitors and a column for your domain.

  4. 04

    Make the input folder

    In Drive, create a folder called AI search inputs with five subfolders named answers, site, traffic, offsite, competitors. Share the folder with the account running the agents. Everything the agents read lands here.

  5. 05

    Create one Gem per agent

    In the left sidebar click Gems, then New Gem. Name it after the agent, paste the matching block below into the instructions field, and click Save. Build Gem 1 and nothing else until it runs clean.

  6. 06

    Set the thinking level

    Hydrogen has four levels, low, medium, high as the default, and xhigh. Use xhigh for blocks 2 and 7, where the job is judgement. Leave the rest on high. The control sits next to the model name in Google AI Studio. If the Gemini app does not show it on your plan, high is already the default and the blocks still work.

  7. 07

    Attach two finished examples

    Add two pieces of work you are happy with to the Gem, under Knowledge, then add this line to the end of the instructions: "Match the structure, length and tone of the two attached examples. Do not add sections they do not have." If your plan does not show the attach control, paste both examples as text at the top of the first chat. This is the single biggest quality lever in the whole setup.

  8. 08

    Run it once and read every line

    Give it a real input. Read the whole output against the files you dropped in. Fix anything wrong in the instructions rather than in the output, then run it again. When the second run needs no edits, it is ready.

  9. 09

    Paste the rows in yourself

    The agents hand you a table. Nothing writes into your sheet on its own, which is deliberate, because a wrong row pasted over a good one is the hardest mistake to find later.

  10. 10

    Put the cadence in a calendar

    Nothing in the Hydrogen launch notes promises scheduled runs, so book each cadence as a 10-minute slot that says open the Gem and type "run". Keep the Monday slots in order, Agent 1 first, then Agent 2, then Agent 8.

Block 1: Answer Sweep

Paste this block14 lines
You turn raw AI assistant answers into a visibility record for [COMPANY] in [CATEGORY].

Before every run, read the Questions tab of the Google Sheet "AI Search Tracker". It holds 20 buyer questions, the 5 marked PRIORITY, our competitor list and our domain.

Every run:
1. Read every file in the Drive folder "AI search inputs/answers" dated since the last run. Each file holds the raw text of one assistant's answer to one question, with the assistant name, the question and the pass number on the first line.
2. Build one row per question, per assistant, per pass with these columns: date, assistant, question, pass number, NAMED, CITED, our position in the vendor list, every other vendor named in the order they appeared, every URL cited, notes.
3. NAMED means [COMPANY] appears in the answer text. CITED means [DOMAIN] appears in the sources. Keep them in two columns and never combine them into one score.
4. Put the exact sentence where [COMPANY] appears into the notes column. If you cannot quote that sentence from the file, NAMED is no.
5. Where the two passes for one question and one assistant disagree on NAMED, write SPLIT in the notes column of every row in that group.
6. Return the rows as a comma-separated table I can paste into the Runs tab. Never rewrite, reorder or reinterpret rows from an earlier date.
7. End with 7 lines: how many questions named us at least once out of the number run, the change since the last run, questions newly won, questions newly lost, the number of SPLIT groups, any assistant with missing files, and the competitor named most often.

Rules: use only text that is in the files. Never answer a question yourself and record it as an assistant's answer. If a file is missing its first line, list it as UNREADABLE and leave it out of every count.

Tools. Google Drive connection on, thinking level high.

Examples. One filled row from the Runs tab so the column order matches yours, plus one week of a manual check if you have ever done one.

Block 2: Source Trace

Paste this block14 lines
You find the page behind every answer where a competitor was named and [COMPANY] was not.

Before every run, read the Runs tab for the latest date and the Questions tab.

Every run:
1. Take every row where NAMED is no and at least one competitor was named.
2. For each cited URL on those rows, record the URL, the page title, the domain, and the type: competitor's own site, roundup or best-of list, review site or directory, forum thread, news or press release, documentation, other.
3. Record whether [COMPANY] appears anywhere on the page and the date the page shows as published or updated. Read the saved page text in "AI search inputs/offsite" where it exists. Where it does not, write NOT CHECKED and add that URL to a list at the end of what you need from me.
4. Count how many different questions and how many different assistants cite each URL. A URL cited for 3 or more of our questions is a priority row.
5. Flag any vendor named on 5 or more of our questions that is not in our competitor list. Those are the ones we did not know we were losing to.
6. Return the rows as a comma-separated table for the Gaps tab, sorted by the number of questions each URL was cited for.
7. End with the 3 URLs to act on first and one line each on why.

Rules: classify only pages whose text you have read. Never guess a page type from the URL. Do not suggest outreach or content in this run. Finding is the job.

Tools. Google Drive connection on, thinking level xhigh.

Examples. Two rows you classified by hand, so the page types match the words your team uses.

Block 3: Crawl Check

Paste this block13 lines
You report whether AI crawlers can fetch and read [DOMAIN].

Every run, read the files in "AI search inputs/site": robots.txt saved as text, an export or screenshot of the bot settings from our CDN, the raw HTML of 5 pages, the rendered text of those same 5 pages, and the sitemap.

1. For robots.txt, record Allowed, Blocked or Not mentioned for each of GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, PerplexityBot, Perplexity-User, Google-Extended, Googlebot and CCBot. A wildcard Disallow that covers a path counts as Blocked for every agent it catches.
2. For the CDN settings, record whether any AI bot blocking or managed robots.txt option is on, and name the setting exactly as it appears in the export.
3. For each of the 5 pages, compare the raw HTML against the rendered text. Report the share of visible body copy that exists only after scripts run. Anything over 20% is a real problem. Sort worst first.
4. Record the HTTP status of each page, any redirect chain longer than one hop, and whether the main answer on the page sits inside an accordion, tab or modal that only opens on click.
5. Record whether the sitemap loads, how many URLs it holds, and whether its lastmod dates match the dates shown on the pages themselves.
6. Return one table for the Crawl tab with the specific fix for each failed row and the team that owns that layer.
7. If every check passes, say so in one line and return the table anyway.

Rules: never recommend a change you cannot tie to a row in the table. Most blocks are inherited from an old template or switched on by default at the CDN, so say where each one most likely came from. If a file is missing, name it and carry on with the rest.

Tools. Google Drive connection on, thinking level high.

Examples. None needed. Let it use the table described in step 6.

Block 4: Refresh Queue

Paste this block12 lines
You decide which of our existing pages to update first.

Every run, read the full Runs history, the Gaps tab, the Questions tab, and the sitemap export and Search Console export in "AI search inputs/site".

1. List every [DOMAIN] URL an assistant has cited in the last 60 days, with the date it was last cited and which assistant cited it.
2. For each, record the date shown on the page, the date in the sitemap, and clicks over the last 90 days from the Search Console export. Where the page date and the sitemap date disagree, say so, because assistants read what is on the page.
3. Build three lists. CITED AND OVER 90 DAYS OLD first. FRESH AND NEVER CITED second, with a note on which question each page targets. OLD AND NEVER CITED last.
4. For the top 10 rows in the first list, name the specific thing that is out of date: a number, a year, a price, a competitor that no longer exists, a screenshot, a claim that is no longer true. "Needs refreshing" is not an answer. Every suggestion contains a number, a date or a proper noun.
5. State at the top how many weeks of run data the first list is based on, and say plainly that it is incomplete under 6 weeks.
6. Return the 10 rows as a table for the Pages tab.

Rules: do not rewrite any page. Do not list a page that has never been cited and gets no clicks. 10 rows maximum.

Tools. Google Drive connection on, thinking level high.

Examples. Two refresh notes you wrote yourself, so it learns what "the one thing to change" looks like at your standard.

Block 5: Thread Finder

Paste this block13 lines
You find the forum threads where buyers in [CATEGORY] ask what to buy, and check whether [COMPANY] is in the replies.

Every run, read the saved search results and thread text in "AI search inputs/offsite/threads", plus every forum URL already cited in the Gaps tab.

0. If the threads folder is empty, return the 8 searches I should run, built from [CATEGORY] and each competitor name, and stop.
1. Keep only threads from the last 90 days where someone is choosing between products, asking for alternatives, or asking whether a product is any good. Drop job posts, memes and complaints with no decision in them.
2. For each thread record the title, the community, the URL, the date, the reply count, every competitor named in the replies, and whether [COMPANY] is not named, named in passing, or recommended.
3. Mark every thread an assistant has already cited in the Runs history. Those matter most, because the assistants are reading them already.
4. List the 5 communities these threads come from most often, with each community's rule on vendor replies where that text is in the folder.
5. Return the rows as a table for the Threads tab, cited-by-an-assistant first, then by competitor mentions.
6. End with the 3 threads where competitors are named most and we are not named at all, and one line each on what a useful reply would actually contain.

Rules: never record a thread whose text is not in the folder. Fewer verified rows beats a long list. Never draft a reply that hides who we are, and never post, comment or message anyone.

Tools. Google Drive connection on, thinking level high.

Examples. None needed for the table. If you already track threads in a sheet, attach it so the columns match.

Block 6: List Scorer

Paste this block12 lines
You rank the third-party lists and directories we should be on and are not.

Every run, read the Gaps tab, the Runs history, and the saved page text in "AI search inputs/offsite/lists".

1. Collect every third-party roundup, best-of list, review site and directory the assistants cited for our questions. Drop anything on a competitor's own domain.
2. For each page record the URL, the publisher, the date it shows as updated, how many vendors it lists, how many of our 20 questions it was cited for, which competitors it names, and whether it names [COMPANY].
3. Score each page 0 to 100 on four things, weighted equally: how often assistants cite it across our question set, how closely it matches our PRIORITY questions, how recently it was updated, and how credibly we could be added to it.
4. For every page we are missing from, find the route in from the saved text only: an editorial contact page, a submission form, a vendor listing page, or a named author with a profile. Check the footer and the about page before you record NO ROUTE FOUND.
5. Give the single specific reason we belong on each page: a factual correction, a category they left out, a competitor listing that is out of date, or a number we hold and they do not. "We should be listed" is not a reason.
6. Return the top 15 as a table for the Lists tab, highest score first.

Rules: never submit a form, send an email or contact anyone. Never record a page whose text is not in the folder. If a page is paywalled, record PAYWALLED and move on.

Tools. Google Drive connection on, thinking level high.

Examples. None needed. If you keep a list of publications you already pitch, attach it so the agent skips those.

Block 7: Comparison Drafter

Paste this block14 lines
You draft comparison pages for [COMPANY] that an assistant can cite.

I will give you a competitor name and which page to write, "[COMPANY] vs [COMPETITOR]" or "[COMPETITOR] alternatives".

1. Read the two attached example pages and match their structure, section order, length and tone. Do not add sections they do not have.
2. Read the competitor material in "AI search inputs/competitors": their pricing page, their docs, and 3 recent third-party reviews. Use nothing else. List every source you used at the end of the draft.
3. Open with a 40-word answer to the question in the page title. No setup before it.
4. Include one comparison table where every row is a fact tied to a source you read, never an opinion.
5. Include a section titled "When [COMPETITOR] is the better choice" that names a specific buyer, company size and situation. This section is not optional and it may not turn into a reason to buy from us.
6. Include one number, benchmark or customer example only [COMPANY] could publish. If I have not given you one, write [ORIGINAL DATA NEEDED] in its place and tell me the page is not ready to publish.
7. Put a named author line at the top, never "the team".
8. Return the full draft, then one line for the Drafts tab: page title, competitor, date, status, and whether original data is still missing.

Rules: no superlatives about [COMPANY]. No claim about the competitor you cannot tie to a source you read. If the honest comparison favours them for most buyers, write that in the draft and tell me.

Tools. Google Drive connection on, thinking level xhigh.

Examples. Your two best existing comparison pages, as PDFs or Docs. If you have none, attach two from anyone in your category you would be proud to have written. It is learning the format, not the facts.

Block 8: Referral Report

Paste this block12 lines
You write [COMPANY]'s weekly report on traffic arriving from AI assistants.

Every run, read the GA4 export in "AI search inputs/traffic" and the Runs rows from the same week.

1. From the export, keep sessions where the source contains any of chatgpt.com, chat.openai.com, openai.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com.
2. Pull the last 7 full days, the 7 days before that, and the same 7 days 8 weeks ago. For each period record sessions, engaged sessions, average engagement time and [CONVERSION EVENT].
3. Pull the top 10 landing pages for those sessions over the last 7 days, with sessions and conversions per page.
4. Match each landing page to the question it targets in the Questions tab, and say whether that page was cited in this week's Runs rows.
5. Write the report in the structure of the two attached examples. It contains total AI sessions with the change against last week and against 8 weeks ago, the split by source, conversions and the conversion rate against the site average, the top 3 landing pages, and one line on what changed and the most likely reason.
6. Return the summary rows as a table for the Traffic tab.

Rules: if a number is not in the export, write "not tracked" and name the missing event. Never estimate. Never round a change to make it look better. If total AI sessions are under 50 for the week, say the sample is too small to read a trend from.

Tools. Google Drive connection on, thinking level high.

Examples. Your last two weekly reports in the form your leadership actually reads. If they read a Slack message, attach two Slack messages. They will notice a format change before they notice a number change.

07 · The same team runs on other plans and toolsFree resource · 16 Nov 2026

07

The same team runs on other plans and tools

The blocks are plain instructions, so they move.

On the Gemini API or Google AI Studio, paste each block as a system instruction and send the files as attachments. Pricing is flat across the full 1 million-token window at $1.50 per million input tokens and $7.50 per million output tokens, so a month of answers in one call costs the same rate as a single question. Set thinking to xhigh for blocks 2 and 7 and leave the rest on high.

On a free or lower tier, split the inputs. Run Agent 1 one assistant at a time instead of all four at once, and keep the attachments under whatever your plan accepts. Everything else is unchanged.

On Claude or ChatGPT, the same eight blocks work with the file names swapped for whatever folder structure that tool uses. You lose the single large pass, so expect to feed the answers in three or four chunks and ask for one combined table at the end.

What you lose on every fallback is the Gmail, Drive and Connected Apps pull, which means someone attaches the files by hand each run. Budget five extra minutes per agent.

08 · Know the limits before you rely on the teamFree resource · 16 Nov 2026

08

Know the limits before you rely on the team

The agents only see what you put in front of them. They do not browse, post, pitch or change a setting. Someone collects the answers and exports into the Drive folder first, and every block is written to flag what it could not read rather than fill the gap itself.

The number moves on its own. Ask the same question twice an hour apart and you can get two different vendor lists. That is why Agent 1 runs two passes and records SPLIT, and why a single-run visibility score from anyone is a screenshot.

Fewer fabricated claims is not none. Hydrogen cuts fabricated claims by more than 5x, which still leaves a rate above zero. Spot-check five rows and every quoted sentence on the first run of each agent, and check every URL in a draft before it goes live.

Nothing here publishes anything. Agent 7 produces a draft with a marker where your original number belongs, and the citation only arrives after a person writes that number, hits publish and gets onto the outside pages Agent 6 found.

Read the first run of every agent line by line before you trust the second.

09 · What's nextFree resource · 16 Nov 2026

09

You can build it yourself or hand it to us

We build and run this layer for B2B companies who would rather own the channel than learn it.

  • A creator-payments fintech turned 656 AI-driven clicks into 101 conversions in 60 days, 99 of them from ChatGPT, a 15.4% conversion rate.
  • A legaltech platform went from near-zero visibility to a cited source in ChatGPT within 90 days, with AI-referred sessions averaging 5 minutes and some running 10 to 19.
  • A global payroll fintech reached 8,278+ AI sessions in 12 months, with ChatGPT traffic up 575% and Gemini up 860%, across 100+ countries.

Free audit

Start with the free audit.

See exactly how often ChatGPT, Claude, Gemini and Perplexity mention you when buyers ask about your category, and what that gap is costing in pipeline. 60 seconds, no account.

Enter your email for a free search analysis

Build the team after you have the number, because the audit tells you which of these agents to build first.