Will Spurlock

AI Visibility strategy (AEO/AIO/GEO) + AI agent builds that turn searches into hot leads🔥 20k+ hours in agentic systems + AI development and Automation in n8n

[Will Spurlock @creativewill02]

This site is a demo.
I built it for FULLER to show what an AI-visibility-first build looks like — and it now outranks the original site in almost every AI search.
Every practice is proven and was working so well I had to tone them back becuase its a DEMO, BABE.
🔗https://t.co/0qbK41piZT

#webdesign #AISearch #aivisibility #MusicMarketing

[Will Spurlock @creativewill02]

Google does not pull "a blog post."

It pulls a passage it can steal in one grab.

Lead-answer H2. Table with comparable cells. Short list. Definition, then a receipt.

The essay that hides the answer in paragraph seven does not get lifted. The slogan page does not get lifted. The logo listicle does not get lifted. Adjectives are what the model already invents.

I do not have a public format ranking from Google. Search Central says the page has to be indexed and snippet-eligible. No special schema. No magic file. After that floor, format is whether the retrieval pass can take a sentence without your brand story.

If a stranger cannot screenshot the H2 plus the next two sentences and leave with a usable answer, the section is fog. Rewrite the lead. Then pick one block.

https://t.co/rs8cdKC79X

[Will Spurlock @creativewill02]

The shortlist happens before your homepage.

G2's March 2026 survey of 1,076 B2B software buyers found 69% chose a different vendor than they first planned after chatbot guidance. The buyer asked a chatbot for names. The model returned two or three. Your hero never loaded.

SaaS teams still treat the homepage as the start of the deal. New hero. New demo CTA. More ad spend pointed at people who already decided. That stack assumes the buyer arrives undecided. They arrive to confirm a name they already heard — or they never come.

Chatbots were the #1 source influencing which vendors made the shortlist (54%). A citation from a review site was the #1 confidence signal in an AI answer (45%). If those surfaces never mention you, the homepage rewrite is late.

I treat this as a citation system. Same product name, same category, same constraints on the review profile, the comparison page, the docs, and a public security page. Then a 12-prompt battery across engines so you can see who gets named and which URL got cited.

You are not losing the deal on the hero. You are losing it in the answer that happens first.

[Will Spurlock @creativewill02]

The homepage should hit before the bio-link habit kicks in.

I built this unofficial The Runarounds demo so the first viewport already has a next tap. Band first. Listen where a thumb already is.

Unofficial. Not a client site. Some paths are framing.

https://t.co/AIN3xUsAYZ

Judge the first screen. Follow if it holds.

[Will Spurlock @creativewill02]

A Q-id is not a citation.

Wikipedia as a domain is one of the pages ChatGPT, Perplexity, and Google AI Overviews already credit. That is not the same as get an article and you get cited.

English Wikipedia wants significant coverage in multiple independent sources. Most small brands will not get a page. Wikidata is an identifier. It is not a byline.

If you cannot pass notability, do not spend a quarter on a stub. Build the entity home on your own site. Align name, category, and geography across Google Business Profile and two directories. Then test "what is [brand]" against "who should I hire."

A missing Wikipedia page is normal. A missing entity home is the miss.

https://t.co/GWb405NajJ

[Will Spurlock @creativewill02]

Your homepage is the wrong pile.

An English homepage can load, rank, and still be useless the second a buyer asks the same question in Spanish. Answer engines pull proof written in the language of the question. Spanish questions pull Spanish pages, Spanish reviews, Spanish entity strings. Your English home sits in the English pile. That is a different stack.

I do not tell owners Google punished the English site. Nothing got deleted. Retrieval missed you. The crawl never saw a second-language URL. The reviews stayed English. The directories that name you stayed English. So the Spanish answer names two competitors who published one real locale page — or it names nobody.

Disappear looks like this in a prompt panel. Same buying question. English run names you. Spanish run cites a directory and a competitor. Your URL is not in the list. The brand did not get weaker. The evidence language stayed fixed.

English rank does not transfer into that second pile. Different index. Different citation list. Winning English Google is a statement about English documents. A Spanish answer is a statement about Spanish documents. If you only test in English, you are auditing the easy half.

Language is the unit, not country. A US shop that never ships overseas still vanishes when a customer who already buys from them asks in Spanish. Location is not language. A Spanish prompt from a US IP is still a Spanish retrieval problem.

Keep the English site. Make it citation-grade. Then stop treating the homepage as proof you exist in every language your customers already speak.

[Will Spurlock @creativewill02]

Graduate a template. Never "the agent."

I want 50+ clean gated sends of this exact template, a low edit rate, rejects that cluster on one bad retrieve instead of random tone misses, a weekly sample where the packet equals the sent body, a frozen scope, and an undo path. Then I take that tool off the gate. I do not take the whole agent off it.

What I will auto-send after that sample: internal ops digests (T0). A fixed appointment reminder SMS whose body is a template plus three CRM fields, after the owner has approved the template itself. A form-received email that cannot mention price, dates the CRM does not own, or legal terms.

What I still will not auto-send just because a flagship model had a good month: first-touch sales email that invents a next step. Any message that includes a number the CRM might have stale. Shared Slack channels. Sequences. Quotes, refunds, payment links.

If your graph only has one send node for everything, you cannot graduate. Split the tools. Take the reminder template off the gate. Leave freeform Gmail send on it.

One customer-facing miss puts the path back behind the gate the same day. That is an IF you can flip, not a workshop.

Human-in-the-loop does not mean a person reviews every retrieve. It means a person reviews the irreversible action — usually the send. Gate the wire, not the research.

The write-up, including the clocks, the card fields, and the graduation checklist:

https://t.co/1mdWnI5Vxy

[Will Spurlock @creativewill02]

The model does not change the deny-list.

Tools you attach do.

Operators keep treating a model upgrade like a permission upgrade. Swap Opus for Flash. Swap GPT for Gemini. The agent is suddenly "safe enough" to hold send-as. It is not. A next-token machine with a fat credential is still a next-token machine with a fat credential. The version string is not a control.

Day-one deny stays the same on every model I ship:

send-as and outbound identity
delete and destructive write
payment write
production database write
unrestricted web fetch
identity-admin and role minting

I have spent 20,000+ hours on agentic systems and 600+ automations built / 500+ live. The miss I see most is not a weak prompt. It is one Gmail scope that can send, one processor key that can refund, one database user that can DELETE, one HTTP tool that will fetch any URL the model invents.

OWASP named this LLM06:2025 Excessive Agency — excessive functionality, excessive permissions, excessive autonomy. A mailbox summarizer that can also send is the textbook case. Changing the model does not detach the send tool.

Read-only on a narrow slice of one system is the default. Everything else is an exception with an owner, an expiry, and a human approval path.

Opus or Flash. Same lock.

[Will Spurlock @creativewill02]

They cite the number they cannot copy.

Restating last year's Ahrefs recap is not original research. The engine already has that pile. A dated table from your shop — n, method, engine names — is a number that disappears if they skip your URL.

I run owner-run prompt panels. Twelve to twenty buyer questions. ChatGPT search, Perplexity, Google AI Overviews. Named, cited, recommended as separate cells. ISO date. Logged-out state. One honest pass.

Or I publish the operations count only I can fill. Last 90 days of inbound reasons. Callbacks on the job type I actually sell. Size-exchange rate on one SKU family.

I do not invent a before/after percentage to make Monday ship. One honest point beats a fake slope. Zero is a measurement. Hiding zero is a press release.

HTML table on a public URL. Lead sentence an engine can steal. Method in the next paragraph.

https://t.co/pzvI4ekwUU

[Will Spurlock @creativewill02]

That Slack ping after send is not a gate. It is an audit of damage.

I keep seeing the same stack on otherwise careful builds. The agent drafts a follow-up that is factually fine. The workflow posts the ops channel: "Sent follow-up to Acme." A human reads it eight minutes later. The follow-up already went to the wrong Acme, or CC'd a competitor, or promised a date nobody owns.

That ping trained the team to feel supervised. The send already happened.

Three words I refuse to let teams blur.

Draft means the agent wrote text, picked a recipient, and stored both. The customer has not seen it.

Notification means Slack or email that says "I sent this." You are reading a receipt.

Approval gate means a blocking wait. Execution holds until approve, reject, edit, or timeout. A thumbs-up on a message that already left is not that.

Notify-after proves someone was told. Approve-before proves a named person released the send. Draft-only is honest if no auto-send is attached. Notify-after is theater.

Outbound is irreversible in a way a chat reply is not. You can edit a support widget. You cannot unsend a thread that already hit a GC's inbox, a prospect's phone, or a CRM sequence that enrolled 400 contacts. A grounded draft can still be the wrong tone, the wrong recipient, or the wrong promise. Accuracy is not permission to send.

I will take a two-minute pause over a two-hour apology every time. If the send node can run without the click, you do not have human-in-the-loop. You have a notification.

[Will Spurlock @creativewill02]

Fill the catalog first. Title, price, identifier, in-stock. Empty product facts are why ChatGPT and Perplexity name a competitor. An empty catalog means an empty recommendation. Fill each product record. Put title, price, identifier, and in-stock on the live product page at once.

ChatGPT, Perplexity, Google AI Mode, and AI Overviews share a need for extractable product facts. They do not share a pipe. ChatGPT blends memory, browse, and merchant feeds. Perplexity cites live pages plus a merchant catalog. Google AI Mode and Overviews both draw on Search plus the Shopping Graph — Mode stays in the conversation, Overviews sit on the results page. Same product record. Different proof and freshness requirements.

Treat them as four surfaces with one catalog, not four SEO projects.

If catalog, feed, proof, or entity is empty, the recommendation stays empty.

I do not split "GEO budget" into four equal piles. I pick the surface that already has a pulse, then make the catalog good enough that the other three can reuse it. DTC with no Merchant Center: enable it and free listings before any blog sprint. ChatGPT never names you and Perplexity sometimes does: third-party corroboration plus the ChatGPT merchant path. Amazon owns every shortlist: identifiers, official PDP density, specialty retailer mentions. You are losing the record, not the ad auction.

Do not invent an "AI revenue" percentage. Most stores cannot isolate it cleanly yet. Score the prompt panel. Overlay Merchant Center AI performance for Google surfaces. Watch whether the cited link is yours.

The full stack, the engine map, and the 30/60/90 sequence:

https://t.co/LqZsPBi7Pg

[Will Spurlock @creativewill02]

Engines do not browse your theme.

They pull records. Then they name a shortlist. Your cinematic hero, the hover animation, the collection grid that ate six weeks — none of that is a catalog. Recommendation is a retrieval problem. If a model cannot load your offer as facts, it names a denser catalog.

A win in 2026 is not a #1 blue link. It is a named shortlist: product, brand, a reason, sometimes a price band, sometimes a merchant. The shopper may click. They may buy later on a brand search. They may never visit you. If you are not named, you are not in the set.

You cannot buy that citation the way you buy a Product Listing Ad. You earn it with data the engine can trust.

Four layers. Most stores I audit fail two.

Catalog: on-page Product schema plus visible attributes so the model has a public API for each SKU. Feed: Merchant Center, ChatGPT's merchant path, Perplexity's merchant program — price, stock, identifiers current at scale. Proof: reviews, roundups, one URL that is not yours that names the SKU. Entity: one brand string on the site, in schema, and in Merchant Center.

Skip the catalog and the engine guesses. Then it prefers Amazon.

I have watched owners celebrate ranking for "{brand} official" while losing every "best for" prompt in the category. That is not a ranking win. That is a recommendation loss.

Theme work can wait. The record cannot.

[Will Spurlock @creativewill02]

The first screen should feel like a threshold, not another bio link.

I built this unofficial $uicideboy$ demo so the grey has a door — listen, tour, G59 energy a cold visitor can actually use.

Unofficial portfolio concept. Not a client announcement. Not a claim of ownership. Some controls may be theater.

https://t.co/xhcGzSgOz3

If the craft lands, follow.

[Will Spurlock @creativewill02]

Stop buying the research questions AI Mode already answered. Keep brand and phone-intent. Put the leftover into pages that can get cited. The budget problem is paying for mid-funnel research the answer already absorbed. Protect brand. Buy urgency. Cut the informational waste now.

Treat AI Mode as a demand-shaping layer that compresses unpaid research, not as a reason to abandon Google Ads. The accounts that stay efficient protect brand and high-intent, cut informational waste, and put money into citation-ready pages so paid dollars close deals that AI Mode already educated.

As of mid-2026, Google has described AI Mode as Search for longer, conversational, multimodal questions, with Gemini 3.5 Flash as the stated default. What Google has not published is a tidy table that says AI Mode changes your CPC by X percent. Treat third-party CPC shock stories as estimates unless they cite Ads Help, earnings language, or account-level data you can verify.

Simple rule I tell founders:

Defend brand. Always, unless you have a rare monopoly with zero conquesting.

Buy urgency. Phone-intent still justifies auctions.

Earn research. Do not rent it forever at rising CPC.

Measure the handoff. AI Overview into AI Mode is a product path Google has publicly described. Your content should survive both the summary and the follow-up.

Paid closes. Citations teach. Teaching without closing wastes organic. Closing without teaching wastes media.

I was SEO certified in 2021. This is the AEO version of a budget meeting: which queries still need a paid click, and which ones you should have been cited in.

https://t.co/7eDC3rc9vX

[Will Spurlock @creativewill02]

A logo list is not a method.

"Best X 2026" pages built from twelve logos, one-line blurbs, and a buy link give an answer engine nothing it cannot already invent. ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews quote the page that already looks like a decision. They skip the page that looks like an affiliate storefront with ten H2s and no scoring.

I keep seeing the same tells on almost every "best X" page I open.

Secret ranking. "#1" with no weights. The model cannot quote a method that is not on the page.

Stale prices and plan names. Last true two years ago. Buyers notice.

One-line items. "Great for teams." That is a vibe, not a claim.

Affiliate gravity. Every heading is a product name plus a cart. The page reads as a catalog.

No "who should skip this." A ranking with no exclusions is an ad.

If anyone in your category could have published your order this afternoon, you are not a source. You are interchangeable copy.

Google's Search Central guidance on generative AI features draws the same split: a generic tips roundup versus a first-hand account only the author could write. That split applies to "best CRM" and "best project tool" pages. Commodity order is not a citation.

The job is not "rank for best X." The job is be the page the model quotes when it builds the shortlist. Fan-out queries will ask "best for small teams" and "what to skip." If your page cannot answer those, a scored competitor will.

A numbered list of logos is not that page.

[Will Spurlock @creativewill02]

Describe the workflow out loud. Then check the graph.

n8n Assistant takes the outcome — spoken or typed — and drops a real canvas. It is a chat agent, not a separate voice app. Dictation is just how speech becomes the prompt.

I have 600+ automations built / 500+ live. The first graph is cheap. A surprise send is not.

Name the trigger, the apps, the fields, the stop, and "ask before publishing." Read every node. Pin one fake payload. Keep write nodes dark until an owner can say the new node list out loud.

If the assistant added a send you did not name, delete that node before you test. The canvas is a draft until you say otherwise.

https://t.co/LTxDPY7nH5

[Will Spurlock @creativewill02]

Ads did not buy you a citation.

Auctions are not answer slots. That is the budget problem under Google AI Mode, and most accounts still treat it like a bid setting.

The mistake I keep seeing: stuffing ads, schema, and "AI Mode SEO" checklists while shipping pages that cannot answer a multi-step question in a form a model can extract. Ranking in AI Mode is citation and passage quality on the same crawl and trust stack as classic Search. Raising Google Ads budget on research terms does not purchase a citation.

I have watched accounts grow spend on informational queries while AI-answer surfaces took the comparison stage. Schema without substance is decoration. AI Mode synthesizes. It does not reward a Business Profile badge by itself, or a generic service page that never states who you serve, where, and what proof you have.

What ranking means for a budget owner:

Citation, not position 1. There is no classic blue-link slot to own inside the synthesis. You want to be a named source with a clear entity.

Passage fitness. Pages that open with a direct answer, then a table or list, get extracted more cleanly than a narrative that buries the take.

Same foundation as Search. If Search Console shows crawl or index issues, AI Mode will not rescue you.

Paid cannot substitute. Ads can capture intent that still clicks. They cannot force Gemini-powered synthesis to call you the best plumber in Austin if your site never proves it.

Paid search still works. The mix of queries that deserve paid coverage has changed. Ads buy auctions. Citations come from crawlable, answer-first pages. Those are different products. Stop treating the bid like a citation.

[Will Spurlock @creativewill02]

Leave write tools off until the agent is boringly correct on read-only work. Then a named human on the write. Never day-one autonomy. One bad Shopify or HubSpot write can erase months of trust. Grant write access only after two to four weeks of checked judgments. Then log misses.

This is the load-bearing rule on every ops agent I ship. Read-only. Then shadow. Then a named human on the write. Limited autonomy only after a clean streak on a narrow, reversible task.

The failure mode I see constantly: an owner buys "AI runs the business," then grants write access to Shopify, QuickBooks, or HubSpot before anyone has checked judgment on read-only work for two to four weeks.

Sequencing, non-negotiable.

Read-only agent. Pulls data. Ranked options. Never mutates live systems.

Shadow mode. Agent proposes the action. A human notes agreement or disagreement. Log every miss.

Human-in-the-loop writes. Agent prepares the change. Named owner approves. Then the write executes.

Limited autonomy. Only after a clean streak on a narrow task — refunds under a threshold, status updates, internal Slack digests. Not payroll. Not live pricing. Not customer promises.

You cannot trust savings numbers from a write-enabled agent until you trust its judgment on read-only work. Jump to writes and those "hours saved" may be hours spent cleaning up bad records.

Working targets before promoting a task: ~90%+ usable on an internal KPI pack. ~85%+ on high-volume ticket tags. ~90%+ on required CRM fields. ~95%+ vs policy on refund suggestions under threshold. Do not lower the bar because a vendor is impatient.

If a vendor insists on write keys in week one, treat that as a sales motion, not an ops design.

I have spent 20,000+ hours architecting agentic systems. The win is fewer payroll hours burned on work a machine should draft and a human should approve.

https://t.co/wuAKt3Sxaw

[Will Spurlock @creativewill02]

The agent runs the Monday pack. You still decide.

Vendors will sell you the opposite. "AI runs the business." Then they hand the model write keys to Shopify, QuickBooks, or HubSpot before anyone has watched it work on a read-only brief for two weeks. That is how you get a wrong refund, a wrong inventory adjustment, or a client email you cannot unsend.

I have spent 20,000+ hours architecting agentic systems. The pattern that survives contact with payroll and inventory is boring on purpose.

An ops agent analyzes clean data in a read-only lane. High-stakes action routes through a human who owns the P&L. Everything else is a demo that fails the first week it touches money or stock.

Keep the rows separate.

Observation: the agent. "Refund rate up 1.2 points week over week."

Recommendation: the agent. "Pause SKU bundle X; support cost exceeds margin."

Decision: you. "Pause the bundle Friday. Update the site Monday."

Execution: the agent only after approve. Named owner. Audit log.

If the vendor blurs those rows, you bought strategy theater. The agent is a tireless analyst that never forgets to run the Monday pack — if you feed it truth and keep a human on the approval button until the read-only miss log is dull.

Three things you name out loud or you do not have an ops agent.

Source of truth. Which system wins when the store, the warehouse sheet, and the CRM disagree?

Freshness. Last night's sync, or a CSV someone emailed three Fridays ago?

Decision owner. Who acts this week, and what budget can they move without another meeting?

If Monday already breaks when one coordinator is out, you do not have an "AI opportunity." You have a single point of failure. Harden it. Do not crown a chatbot strategist.

[Will Spurlock @creativewill02]

Airtable AI is enough when the job stays in one base.

Omni. Field Agents. Generate with AI. Drafts you will actually read.

You need an outside agent when two systems have to join on a clock you do not sit in.

That is n8n. Not a second chat window.

Stay inside Airtable when the record, the instruction, and the output live in the same table. Leave when the run has to read Gmail, write a sheet, and ping Slack without you opening the base.

I do not buy a seventh app to summarize a view I already pay for. I also do not pretend Field Agents can replace a graph that crosses vendors.

The test I use: if I close Airtable, does the job still need to finish? If yes, wire n8n. If no, flip the in-base draft and keep send off.

https://t.co/FSy40RO9MN

[Will Spurlock @creativewill02]

Two hours. Six apps. One draft switch.

If it takes a week, you are shopping again.

List the six apps you already pay for. Skip any row you do not pay for. Do not install a tool to finish the table.

Turn on one draft-only feature. Leave send off.

Then stop shopping this week.

Open each vendor's AI settings page. Match the feature to the plan you actually have, not the plan the demo assumed. Slack summaries can sit on Pro. Recaps and search answers sit on Business+ and Enterprise+. Notion AI is Business and Enterprise. HubSpot Assistant is included; Agent Hub is a different card.

Edit the bad ones. Tighten the instruction. Leave money fields off. Leave web search off any price column.

The 120-minute pass I actually run: minutes 0–15 confirm who can flip an admin lock. 15–35 Notion, one Basic Autofill on a database you already live in. 35–55 Airtable, one field agent, check the credit balance. 55–75 HubSpot, Breeze Assistant on one live deal — do not Activate five Agent Hub cards. 75–95 Slack recap or Gemini in Gmail. 95–120 Shopify Magic on one unpublished product, Sidekick on Review only.

Friday is good when you can name the one feature you left on, you have ten before/after examples even if they are ugly, and you have not granted a new vendor write-access to email. If four rows are Off and you were about to buy an agent, you do not have an agent problem. You have unused inventory.