bgj

building with claude code

[bgj @benjamingjoel_]

nigeria’s biggest problem isn’t only corrupt politicians. it’s pastors who convince the poor that their poverty is god’s plan, so nobody ever demands better.

[bgj @benjamingjoel_]

@prophetbenomale @prophetbenomale you just wasted your time to write rubbish..

so much spiritual power, and nigeria still has no light, no security and no jobs. either the power is fake, or you're wasting it on x fights.

[bgj @benjamingjoel_]

@soccertorey @NWA_ADA_NNEWI Bastoni

[bgj @benjamingjoel_]

content behind a "load more" button is content the model skips.

if your answer only appears after a click, a scroll-triggered load, or a "read more" expand, assume the fetcher never sees it.

1/ it grabs what's in the initial response. it does not click your button.
2/ so faqs that expand on tap, comments that lazy-load, specs hidden in an accordion, all invisible.
3/ anything you actually want quoted has to be there before any interaction. flat, in the html, on arrival.

watched a genuinely good faq section score zero because every answer was one tap away. the tap was the wall.

[bgj @benjamingjoel_]

your product name might be tokenizing weird and the model mis-reads it.

if the name is camelCase, has a special char, or is a normal english word, the retriever can split it or confuse it with the common meaning. saw a model merge a client's product with an unrelated word that happened to be their name.

cheap fix: pair the name with the category every time it matters. "Acme, the invoicing tool" not just "Acme". gives the model an anchor so it stops guessing which acme you are. did it on their top 15 pages.

[bgj @benjamingjoel_]

your cache-control headers might be blocking your own re-crawl.
client whose updated pages weren't showing new info in answers for weeks. wasn't the model. was them.
1/ their cdn set cache-control max-age to a year on html
2/ so fetchers cached the old version and saw no reason to come back
3/ dropped max-age to something sane, pages got re-fetched, answers updated within the week
you can serve a perfect page and still tell the crawler "nothing to see here for 12 months". check the header, not just the page.

[bgj @benjamingjoel_]

the fetcher is probably reading your mobile build, not your desktop one.

this trips people up. you look at your gorgeous desktop page with all the detail and assume that's what the model sees. a lot of setups serve a stripped mobile or amp version to bots, and that version quietly drops half the copy: the comparison table, the faq, the specifics.

so the version you're proud of and the version getting read are two different pages. desktop convinces a human, the mobile skeleton is what gets quoted.

worth checking which one your fetcher actually gets. we scan for this if it's useful, https://t.co/69Qi6Lp5zI.

[bgj @benjamingjoel_]

your page title is the headline the model uses, and yours probably wastes it.

"Acme | Home" or "Acme | Enterprise Data Solutions" tells a model nothing about what you actually answer. the title is prime real estate and half the web spends it on the brand name.

we rewrote ours to say what the page actually answers, in the words people ask. small change, the model started framing us more accurately. put the answer in the title, not just the logo

[bgj @benjamingjoel_]

schema markup, old head vs where i've landed.

old head: schema is for google. get the rich snippet, the star ratings, the little price tag in search.

where i am now: schema is how you hand a model clean facts it otherwise has to guess at. price, what the thing is, who makes it, how it relates to your other stuff.

instead of hoping the bot parses your pretty layout correctly, you give it the structured version and take out the guessing.

we mark up product facts this way now and the models quote them back more accurately. the patterns on https://t.co/69Qi6Lp5zI if you want to see it. way less "the model got our price wrong", because we stopped making it infer the price from a styled div

[bgj @benjamingjoel_]

how to write a page so the model can grab the exact answer instead of the whole thing:

1/ put the actual question as an h2, in the words people type. "does it work offline" not "connectivity"
2/ answer it in the first sentence under that. dont warm up
3/ one question per section. a heading answering three things matches none of them
4/ stick an id on each heading so the bot can grab that chunk directly
5/ keep each answer able to stand alone. no "as mentioned above", they quote out of order

we do product pages like this now, the patterns on https://t.co/PbuMcMj9OO. same words, way more of it gets quoted

[bgj @benjamingjoel_]

stuff that makes your best page invisible to a model even though it ranks fine for people:
content thats behind a click. tabs, accordions, "read more"
text baked into an image instead of real html
the answer sitting 2000 words down with no heading on it
infinite scroll that never gives the bot a real url
a cookie wall that loads before the actual page does
none of these really hurt you on google anymore. all of them can make you unquotable. spend an afternoon crawling your own site like a bot would, its grim

[bgj @benjamingjoel_]

ok this bugged me today. chatgpt is citing our changelog and not the landing page. the one we paid someone to write.
losing to a list of "fixed x, shipped y".
think its just the dates + it says literally what the thing does. landing page is all "transform your workflow" and the model cant do anything with that. anyway dont delete your changelog

[bgj @benjamingjoel_]

ok this bugged me today. chatgpt is citing our changelog and not the landing page. the one we paid someone to write.
losing to a list of "fixed x, shipped y".
think its just the dates + it says literally what the thing does. landing page is all "transform your workflow" and the model cant do anything with that. anyway dont delete your changelog

[bgj @benjamingjoel_]

EU AI Act says mark your AI content. anthropic's watermarking claude's text. apple's marking iphone photos as human-made.

so does this video look AI-generated to you, or what?

[bgj @benjamingjoel_]

@boardyai yes please, would love an intro to both

[bgj @benjamingjoel_]

double-check your robots.txt before you blame your content.

a lot of sites pasted in a "block the ai scrapers" snippet last year and forgot. now they're fully indexed in google and completely invisible to chatgpt, because they told gptbot to go away in one line nobody remembers writing.

[bgj @benjamingjoel_]

your best answers are probably locked behind a login.

support macros, the knowledge base, the detailed how-tos, all the stuff that actually answers buyer questions, sitting inside a gated help center the model can never reach.

meanwhile it answers those exact questions from some random forum thread. move the evergreen, non-sensitive articles onto a public, crawlable docs page. that's where your real answers earn citations.

[bgj @benjamingjoel_]

if a page is thin, the model will sometimes just lift your meta description as the summary of you.

which is funny, because most meta descriptions are an afterthought some plugin generated. a vague 150-character blurb becomes the sentence a buyer reads about what you do.

worth writing the important ones by hand, as a real answer to "what is this," not seo filler. small surface, punches way above its size.

[Simon Wilhelm @Simon_LeanderW]

your clearest explanation is probably in a video the model can't watch.

1/ a fetcher reads text. a 12-minute youtube embed or a wistia player is a black box to it, no matter how good the content is.
2/ so if your best answer only exists as speech, it's invisible to every ai engine.
3/ drop a full transcript on the page as real text, not a downloadable file. now the model can read what you actually said.

we do this for our own demos. the transcript is what ends up getting pulled into answers.

free scan of what's actually readable on your site: https://t.co/n2qoUMeAoA

[bgj @benjamingjoel_]

Ever asked ChatGPT whether you should take a job somewhere?

Your candidates have. Before they answer a recruiter, plenty of them ask the model what your company is like to work at, whether you're stable, whether the product is any good.

It answers from Glassdoor, an old layoff headline, a Reddit thread. If your story out there is thin or stale, your offer competes with whatever it scraped. Might be worth reading that answer before your next good candidate does.